CN117830932A - Obstacle detection method based on continuous frame point cloud space feature enhancement - Google Patents
Obstacle detection method based on continuous frame point cloud space feature enhancement Download PDFInfo
- Publication number
- CN117830932A CN117830932A CN202311703585.9A CN202311703585A CN117830932A CN 117830932 A CN117830932 A CN 117830932A CN 202311703585 A CN202311703585 A CN 202311703585A CN 117830932 A CN117830932 A CN 117830932A
- Authority
- CN
- China
- Prior art keywords
- point cloud
- point
- points
- clustering
- sub
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 40
- 238000000034 method Methods 0.000 claims abstract description 94
- 239000011159 matrix material Substances 0.000 claims abstract description 49
- 230000009466 transformation Effects 0.000 claims abstract description 37
- 238000012545 processing Methods 0.000 claims abstract description 6
- 238000009826 distribution Methods 0.000 claims description 23
- 230000001133 acceleration Effects 0.000 claims description 18
- 230000006870 function Effects 0.000 claims description 18
- 230000004927 fusion Effects 0.000 claims description 18
- 230000008569 process Effects 0.000 claims description 16
- 230000011218 segmentation Effects 0.000 claims description 14
- 238000004364 calculation method Methods 0.000 claims description 9
- 238000013507 mapping Methods 0.000 claims description 9
- 238000013519 translation Methods 0.000 claims description 7
- 238000001914 filtration Methods 0.000 claims description 6
- 230000003068 static effect Effects 0.000 claims description 6
- 230000008859 change Effects 0.000 claims description 4
- 238000000605 extraction Methods 0.000 claims description 4
- 238000000354 decomposition reaction Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 230000009467 reduction Effects 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 238000009827 uniform distribution Methods 0.000 claims description 3
- 206010034719 Personality change Diseases 0.000 claims description 2
- 241001632422 Radiola linoides Species 0.000 claims 1
- 230000008447 perception Effects 0.000 abstract description 5
- 238000009825 accumulation Methods 0.000 abstract description 3
- 238000005259 measurement Methods 0.000 abstract description 3
- 230000003044 adaptive effect Effects 0.000 description 9
- 238000004458 analytical method Methods 0.000 description 7
- 230000000694 effects Effects 0.000 description 4
- 238000007781 pre-processing Methods 0.000 description 4
- 230000009471 action Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000007613 environmental effect Effects 0.000 description 3
- 241001632427 Radiola Species 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 239000003086 colorant Substances 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
- G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an obstacle detection method based on continuous frame point cloud space feature enhancement. According to the invention, the spatial density of a target object is effectively improved, the spatial characteristics of point cloud are enhanced, the accuracy of target detection is improved, the robustness of inertial navigation measurement data is increased, the method has important application value to the unmanned perception field, and aiming at the problems of offset error, weak robustness and the like of an IMU estimated posture transformation matrix, the precise registration of the IMU estimated posture transformation is provided by using an NDT registration algorithm, so that the error caused by integral accumulation of the IMU is reduced. And then extracting a dynamic target by using a traditional target detection algorithm, independently processing and fusing, and projecting the background point cloud of the previous frame to the point cloud of the current frame according to the estimated gesture transformation matrix, so that the density of the sparse point cloud is improved, and the accuracy of obstacle detection is improved.
Description
Technical Field
The invention belongs to the technical field of environment perception, and particularly relates to an obstacle detection method based on continuous frame point cloud space characteristic enhancement.
Background
The environment awareness technology is a core technology in an unmanned system and is also a hot spot for long-term research in academia. The laser radar is widely applied to an automatic driving sensing system by virtue of the advantages of high precision, high resolution, long distance, strong anti-interference capability and the like. The surrounding environment point cloud data collected by the laser radar can provide accurate geometric information and depth information of an object, and the collected environment information is more abundant along with the fact that the laser beams are more and denser, so that the tasks such as target detection and tracking are facilitated. However, with the increase of the wire harness, the cost is also continuously increased, the price of the high-wire-harness laser radar is often more than ten times of that of the low-wire-harness radar, and the popularization of the automatic driving vehicle is also hindered by the high price of the laser radar, so that how to use the low-wire-harness radar to realize better environment perception has great significance to the automatic driving field,
aiming at the problem of sparse point cloud, liu et al propose a target detection tracking deep learning module based on cameras, laser radars and millimeter wave radars, which integrates the data characteristics of different sensors into a common representation, and the algorithm can accurately identify and track obstacles. Zhang Yanguo et al estimate the vehicle pose by using an inertial measurement unit, further estimate the pose transformation matrix by pose transformation between two frames of point clouds, and directly fuse the two frames of point clouds through the transformation matrix, so that the density of the point clouds can be effectively enhanced.
However, only the pose estimation of the IMU is insufficient in robustness, so that the problems of offset error, low robustness and the like of an estimated pose transformation matrix are caused, and the accuracy of obstacle detection is not high enough.
Disclosure of Invention
The invention aims at: in order to solve the problems, an obstacle detection method based on continuous frame point cloud space feature enhancement is provided.
The technical scheme adopted by the invention is as follows: an obstacle detection method based on continuous frame point cloud space feature enhancement is characterized by comprising the following steps of: the detection method based on continuous frame point cloud space feature enhancement comprises the following steps: s1: the point cloud preprocessing is firstly carried out and mainly comprises three parts of downsampling, outlier removal and ROI (Region of Interest) area extraction. In order to accelerate the operation speed and reduce the storage consumption, the point cloud of the dense area needs to be downsampled, and the characteristic space is ensured to be unchanged;
s2: ground segmentation is carried out, ground information and space information are separated, and subsequent point cloud clustering and obstacle recognition are facilitated;
s3: extracting a dynamic target, namely extracting the dynamic target by using a DBSCAN (Density-Based Clustering Algorithm with Noise, DBSCAN) clustering algorithm;
s4: and carrying out dynamic target point cloud fusion, and after extracting dynamic targets of two adjacent frames of point clouds, carrying out data association matching on the dynamic targets in the two frames of point clouds. The method adopts a nearest neighbor data association (Nearest Neighbor Data Association, NNDA) method to realize the matching between moving objects between two frames of point clouds. The method comprises calculating an incidence matrix according to Euclidean distance between obstacles, taking maximum value of each row in the incidence matrix as incidence result of the obstacles, assuming N objects exist in clustering result of the previous frame point cloud, and M objects exist in clustering result of the current frame point cloud, the incidence matrix can be expressed as formula, wherein D is as follows i,j And representing the Euclidean distance between the ith object in the point cloud of the previous frame and the jth object in the point cloud of the current frame.
S5, carrying out static background point cloud fusion, acquiring three-axis acceleration and attitude angle of the vehicle by using an inertial navigation system of CGI-410 of China survey, estimating an attitude transformation matrix between continuous frame point clouds by using an NDT algorithm, and estimating an initial transformation matrix by using the three-axis acceleration and the attitude angle provided by the inertial navigation system, thereby eliminating the problem of weak robustness in the registration process. Assume that the inertial navigation system is at t 1 The three-axis acceleration obtained at the moment is a 1 Three-axis speed v 1 Attitude angle θ 1 At t 2 The acceleration obtained at the moment is a 2 Attitude angle θ 2 Wherein each variable contains components of the x, y, z axes. The acceleration added in the process is deltaa, the changed attitude angle is deltatheta, and the method can be usedAnd calculating a rotation matrix R of the vehicle posture change at two moments according to the formula, and calculating a translation vector t according to the formula, wherein the finally solved initial transformation matrix init_trans is shown in the formula.
S6, firstly rasterizing the point clouds by using an NDT algorithm, and representing the point cloud distribution in each grid by using a normal distribution probability function, wherein the normal distribution probability density function is calculated in a mode shown in the formula, and the normal distribution probability function is calculated by using a standard distribution probability functionRepresenting a normal distribution probability density function, +.>Represents the mean vector, Σ represents the covariance matrix, +.>Representing a point cloud within a grid, c 1 Is constant.
After the probability density function is obtained, mapping the point cloud to the target point cloud through initialized transformation parameters, determining normal distribution corresponding to each mapping point, and summing the probability density of the mapping points to obtain a score value s (p) of each coordinate transformation parameter, wherein the score value s (p) is shown as a formulaRepresenting a point cloud vector within the mapped grid, +.>Representing the mapped mean vector. And then, obtaining the optimal estimation of the transformation parameters by using a Hessian matrix method and a minimum value of s (p), and finally obtaining the optimal estimation of the transformation parameters until the optimization function converges by using a Newton iteration method.
After the optimal transformation matrix is obtained through the NDT, the point cloud of the previous frame can be projected to the point cloud of the current frame through rotation and translation, so that fusion of the background point cloud is realized.
In a preferred embodiment, in the step S, the method can quickly and effectively dilute the point cloud, and is widely applied to three-dimensional point cloud processing. The method comprises the steps of firstly establishing an axial bounding box for a three-dimensional point cloud, dividing the bounding box into n equal parts along each coordinate axis direction to obtain n multiplied by n voxels, selecting the mass center of the voxel from each voxel to represent the voxel, and completing voxel grid downsampling, wherein the mass center selecting method is shown as the formula, wherein (x c ,y c ,z c ) Represents centroid coordinates, n represents the total number of voxel midpoints, (x) i ,y i ,z i ) Representing the coordinates of the i-th point in the voxel. The voxel grid method downsampling has the advantages of high calculation efficiency, uniform distribution of sampling points, effective reduction of data volume and acceleration of operation speed.
In a preferred embodiment, in the step S1, filtering of the point cloud outliers is implemented by using a method of point cloud radius filtering. The basic idea of the method is to assume that each data point in the initial point cloud has at least a specified number of neighborhood points in a specified radius neighborhood, the points meeting the conditions are regarded as normal point reservation, and the rest points are regarded as outlier removal. For the ith point p in the point cloud containing N points i Calculated as p i Is the sphere center, r isThe number n of points in the sphere of radius, if the number of points in the neighborhood is smaller than a certain threshold T r If the point is determined to be an outlier, otherwise, the normal point is regarded as reserved, and the specific calculation is shown in the formula. The method sets T r For 20, the point cloud data is stored by using a KD Tree (K-Dimensional Tree) data structure so as to realize quick search of the neighborhood points.
According to the characteristic of near-dense and far-sparse of the collected point cloud, the method sets the extracted point cloud ROI area to be [ -70m,70m ] along the x-axis, 20m ] along the y-axis and limits the height along the z-axis to be 2.5m. And extracting indexes of points in the ROI by using the established KD tree data structure, so that the segmentation of the ROI is completed, and the speed of point cloud segmentation is increased.
In a preferred embodiment, in the step S2, since the real ground tends to have many slopes, which is not a perfect plane, it is necessary to deal with the situation that there is a slope change in the ground in order to complete the ground segmentation well. The method utilizes an algorithm based on a plane model to fit the ground, the idea of the algorithm is that firstly, a space is divided into N sub-planes along the x-axis direction (the head direction) of a point cloud, the points in the ith sub-plane are firstly ordered according to the size of a z-axis, l points in the ith sub-plane are selected to form a seed point set, and the plane normal vector of the ith sub-plane is estimated through singular value decompositionThereby obtaining the sub-plane fitting ground sigma i . Then, carrying out orthogonal projection judgment on each point in the ith sub-plane by using an iterative method, wherein the specific method is that if the point reaches the estimated ground sigma i Is smaller than threshold th d Then consider the point to be a ground point and add the point to sigma i Otherwise, do not join, thereby iterating N iter Fitting ground sigma of ith sub-plane can be obtained i Finally, the ground fitted by each sub-plane is spliced together to form the finalThe ground plane σ is estimated, where the fit plane may be represented as shown in the equation, where d represents the intercept, and the final estimated ground may be represented as shown in the equation.
σ＝[σ 1 ,L,σ N ] (4)。
In a preferred embodiment, in step S3, in order to accelerate the clustering process, the point cloud is often projected to a horizontal plane for clustering, so as to reduce the amount of computation in the clustering process. Because the DBSCAN algorithm is sensitive to the neighborhood radius and the minimum point threshold, the method firstly divides the point cloud into M sub-point clouds according to the characteristics of near-dense and far-sparse point clouds, constructs KD tree data structures in each sub-point cloud to improve the efficiency of a clustering algorithm, and sets different clustering radiuses and the minimum point threshold according to the difference of the distances from the center of each sub-point cloud to the origin of the laser radar so as to realize self-adaptive DBSCAN clustering, wherein the clustering radiuses and the distances from the sub-point clouds to the origin of the laser radar form a negative correlation relation, and a specific relational expression is shown in a formula (x) j ,x j ) Represents the jth point, epsilon, in the ith sub-point cloud i Representing the clustering neighborhood radius of the ith sub-point cloud, S i Representing the ith sub-point cloud.
A flowchart of the adaptive clustering algorithm performed in each sub-point cloud is shown in fig. 2.
In a preferred embodiment, in the step S3, for the input segmented sub-point cloud, all points in the point cloud are marked as unvisited points (unvisited), and then the storage structure Q of the point cloud is constructed by using the KD-tree to facilitate subsequent operations. And randomly taking a point P from the point cloud, and judging whether the point P is a core point or not by calculating the distance between the point P and the neighborhood point. If the point is a core point, the point can be mutually communicated with other core points, a new cluster is created, the point is added into the cluster, all points in the neighborhood of the point are used as a seed point set, the core points are searched in the seed point set in an iterative mode, namely, whether the point is the core point or not is judged by randomly taking one point in the seed point set, if yes, the neighborhood point of the point is used as the seed point, and if not, the point is marked as the visited point (visual). And iterating until all seed points are accessed, and outputting a set C for finally storing the clustering result if no point which is not accessed exists in the point cloud Q.
In a preferred embodiment, in the step S3, in order to prevent the clustering of the objects with too small or too large sizes, the clustering needs to be limited, and the method sets a threshold interval of the clustering cluster points, and when the clustering result points are not in the threshold interval, the clusters are not reserved.
In a preferred embodiment, in the step S4, the dynamic target of the matching in the two-frame point cloud may be determined according to the maximum value of each row of the correlation matrix by calculating the correlation matrix, however, such matching may cause mismatching between the adjacent pedestrians and vehicles, and in order to avoid such a problem, some limitation conditions need to be introduced. The most common constraint is that according to the size of the 3D bounding box of two frames of clustering targets, the method suppresses mismatching by judging the relation between the ratio of the volumes of the 3D bounding boxes of the matching targets and a set threshold, and only the targets meeting the formula and the matching result of the incidence matrix simultaneously are considered to be successful in matching, wherein the specific calculation mode is shown as the formula, and l is as follows i ，w i ，h i Representing the length, width and height of the ith object of the point cloud of the previous frame, T v Representing a volume ratio limiting threshold. After the matching of the dynamic targets is completed, the distance between the centers of the matched dynamic targets is calculated, and the dynamic targets in the point cloud of the previous frame are translated to the corresponding positions of the current point cloud according to the distance.
In summary, due to the adoption of the technical scheme, the beneficial effects of the invention are as follows:
according to the invention, the spatial density of a target object is effectively improved, the spatial characteristics of point cloud are enhanced, the accuracy of target detection is improved, the robustness of inertia is increased, the method has important application value to the unmanned perception field, and aiming at the problems of offset error, weak robustness and the like of an IMU estimated posture transformation matrix, the method is provided for carrying out fine registration on the IMU estimated posture transformation by using an NDT (Normal Distribution Transformer, NDT) registration algorithm, so that the error caused by integral accumulation of the IMU is reduced. And then extracting a dynamic target by using a traditional target detection algorithm, independently processing and fusing, and projecting the background point cloud of the previous frame to the point cloud of the current frame according to the estimated gesture transformation matrix, so that the density of the sparse point cloud is improved, and the accuracy of obstacle detection is improved.
Drawings
FIG. 1 is a flowchart of a continuous frame point cloud spatial feature enhancement algorithm of the present invention;
FIG. 2 is a flowchart of a neutron point cloud adaptive clustering algorithm in the present invention;
FIG. 3 is a graph of NDT point cloud registration performance results in the present invention;
FIG. 4 is a graph of the result of cloud preprocessing in the present invention;
FIG. 5 is a graph of the ground segmentation result in the present invention;
FIG. 6 is a graph of the point cloud adaptive clustering result of the present invention;
FIG. 7 is a graph of the result of NDT point cloud registration in the present invention;
FIG. 8 is a graph of the dynamic target point cloud fusion results in the present invention;
FIG. 9 is a graph of the result of static background point cloud fusion in the present invention;
fig. 10 is a graph of the detection result of the point cloud after the spatial feature enhancement in the present invention.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
With reference to figures 1-10 of the drawings,
examples:
the detection method based on the continuous frame point cloud space feature enhancement comprises the following steps:
and S1, performing point cloud pretreatment, wherein the pretreatment of the point cloud mainly comprises three parts of downsampling, outlier removal and ROI (Region of Interest) area extraction. In order to accelerate the operation speed and reduce the storage consumption, the point cloud of the dense area needs to be downsampled, and the characteristic space is ensured to be unchanged, and the method adopts a downsampling method based on a voxel grid method
S2, ground segmentation is carried out to separate ground information and space information, so that subsequent point cloud clustering and obstacle recognition are facilitated
S3, extracting dynamic targets, namely, extracting the dynamic targets by utilizing a DBSCAN (Density-Based Clustering Algorithm with Noise, DBSCAN) clustering algorithm
S4, carrying out dynamic target point cloud fusion, and after extracting dynamic targets of two adjacent frames of point clouds, carrying out data association matching on the dynamic targets in the two frames of point clouds. The method adopts a nearest neighbor data association (Nearest Neighbor Data Association, NNDA) method to realize the matching between moving objects between two frames of point clouds. The method comprises calculating an incidence matrix according to Euclidean distance between obstacles, taking maximum value of each row in the incidence matrix as incidence result of the obstacles, assuming N objects exist in clustering result of the previous frame point cloud, and M objects exist in clustering result of the current frame point cloud, the incidence matrix can be expressed as formula, wherein D is as follows i,j And representing the Euclidean distance between the ith object in the point cloud of the previous frame and the jth object in the point cloud of the current frame.
S5, carrying out static background point cloud fusion, acquiring three-axis acceleration and attitude angle of the vehicle by using an inertial navigation system of CGI-410 type of Hua-ji, and estimating the attitude between the continuous frame point clouds by adopting an NDT algorithmThe state transformation matrix is estimated by utilizing the triaxial acceleration and the attitude angle provided by the inertial navigation system, so that the problem of weak robustness in the registration process is solved. Assume that the inertial navigation system is at t 1 The three-axis acceleration obtained at the moment is a 1 Three-axis speed v 1 Attitude angle θ 1 At t 2 The acceleration obtained at the moment is a 2 Attitude angle θ 2 Wherein each variable contains components of the x, y, z axes. The acceleration increased in the process is deltaa, the changed attitude angle is deltatheta, the rotation matrix R of the vehicle attitude change at two moments can be calculated according to the formula, the translation vector t is calculated through the formula, and the finally solved initial transformation matrix init_trans is shown as the formula.
S6, firstly rasterizing the point clouds by using an NDT algorithm, and representing the point cloud distribution in each grid by using a normal distribution probability function, wherein the normal distribution probability density function is calculated in a mode shown in the formula, and the normal distribution probability function is calculated by using a standard distribution probability functionRepresenting a normal distribution probability density function, +.>Represents the mean vector, Σ represents the covariance matrix, +.>Representing a point cloud within a grid, c 1 Is constant.
After the probability density function is obtained, mapping the point cloud to the target point cloud through initialized transformation parameters, determining normal distribution corresponding to each mapping point, and summing the probability density of the mapping points to obtain a score value s (p) of each coordinate transformation parameter, wherein the score value s (p) is shown as a formulaRepresenting a point cloud vector within the mapped grid, +.>Representing the mapped mean vector. And then, obtaining the optimal estimation of the transformation parameters by using a Hessian matrix method and a minimum value of s (p), and finally obtaining the optimal estimation of the transformation parameters until the optimization function converges by using a Newton iteration method.
After the optimal transformation matrix is obtained through the NDT, the point cloud of the previous frame can be projected to the point cloud of the current frame through rotation and translation, so that fusion of the background point cloud is realized.
In the step S, the method can be used for quickly and effectively thinning the point cloud, and is widely applied to three-dimensional point cloud processing. The method comprises the steps of firstly establishing an axial bounding box for a three-dimensional point cloud, dividing the bounding box into n equal parts along each coordinate axis direction to obtain n multiplied by n voxels, selecting the mass center of the voxel from each voxel to represent the voxel, and completing voxel grid downsampling, wherein the mass center selecting method is shown as the formula, wherein (x c ,y c ,z c ) Represents centroid coordinates, n represents the total number of voxel midpoints, (x) i ,y i ,z i ) Representing the coordinates of the i-th point in the voxel. The voxel grid method downsampling has the advantages of high calculation efficiency, uniform distribution of sampling points, effective reduction of data volume and acceleration of operation speed.
In the step S1, filtering of the point cloud outliers is implemented by using a method of point cloud radius filtering. The basic idea of the method is to assume that each data point in the initial point cloud has at least a specified number of neighborhood points in a specified radius neighborhood, the points meeting the conditions are regarded as normal point reservation, and the rest points are regarded as outlier removal. For the ith point p in the point cloud containing N points i Calculated as p i The sphere center, r is the number n of points in the sphere with radius, if the number of points in the neighborhood is smaller than a certain threshold T r If the point is determined to be an outlier, otherwise, the normal point is regarded as reserved, and the specific calculation is shown in the formula. The method sets T r For 20, the point cloud data is stored by using a KD Tree (K-Dimensional Tree) data structure so as to realize quick search of the neighborhood points.
According to the characteristic of near-dense and far-sparse of the collected point cloud, the method sets the extracted point cloud ROI area to be [ -70m,70m ] along the x-axis, 20m ] along the y-axis and limits the height along the z-axis to be 2.5m. And extracting indexes of points in the ROI by using the established KD tree data structure, so that the segmentation of the ROI is completed, and the speed of point cloud segmentation is increased.
In the step S2, since the real ground often has many slopes, which is not a perfect plane, the ground with a certain slope change must be processed to complete the ground segmentation well. The method utilizes an algorithm based on a plane model to fit the ground, the idea of the algorithm is that firstly, a space is divided into N sub-planes along the x-axis direction (the head direction) of a point cloud, the points in the ith sub-plane are firstly ordered according to the size of a z-axis, l points in the ith sub-plane are selected to form a seed point set, and the plane normal vector of the ith sub-plane is estimated through singular value decompositionThereby obtaining the sub-plane fitting ground sigma i . Then, carrying out orthogonal projection judgment on each point in the ith sub-plane by using an iterative method, wherein the specific method is that if the point reaches the estimated ground sigma i Is smaller than threshold th d Then consider the point to be a ground point and add the point to sigma i Otherwise, do not join, thereby iterating N iter Fitting ground sigma of ith sub-plane can be obtained i And finally, splicing the ground fitted by each sub-plane together to form a final estimated ground plane sigma, wherein the fitted plane can be represented as shown in a formula, d represents an intercept, and the final estimated ground can be represented as shown in the formula.
σ＝[σ 1 ,L,σ N ] (4)。
In step S3, in order to accelerate the clustering process, the point cloud is often projected to a horizontal plane to perform clustering, so as to reduce the operand in the clustering process. Because the DBSCAN algorithm is sensitive to the neighborhood radius and the minimum point threshold, the method firstly divides the point cloud into M sub-point clouds according to the characteristics of near-dense and far-sparse point clouds, constructs KD tree data structures in each sub-point cloud to improve the efficiency of a clustering algorithm, and sets different clustering radiuses and the minimum point threshold according to the difference of the distances from the center of each sub-point cloud to the origin of the laser radar so as to realize self-adaptive DBSCAN clustering, wherein the clustering radiuses and the distances from the sub-point clouds to the origin of the laser radar form a negative correlation relation, and a specific relational expression is shown in a formula (x) j ,x j ) Represents the jth point, epsilon, in the ith sub-point cloud i Representing the clustering neighborhood radius of the ith sub-point cloud, S i Representing the ith sub-point cloud.
A flowchart of the adaptive clustering algorithm performed in each sub-point cloud is shown in fig. 2. .
In the step S3, for the input segmented sub-point cloud, all points in the point cloud are marked as unvisited points (unvisited), and then the storage structure Q of the point cloud is constructed by using the KD-tree, so as to facilitate subsequent operations. And randomly taking a point P from the point cloud, and judging whether the point P is a core point or not by calculating the distance between the point P and the neighborhood point. If the point is a core point, the point can be mutually communicated with other core points, a new cluster is created, the point is added into the cluster, all points in the neighborhood of the point are used as a seed point set, the core points are searched in the seed point set in an iterative mode, namely, whether the point is the core point or not is judged by randomly taking one point in the seed point set, if yes, the neighborhood point of the point is used as the seed point, and if not, the point is marked as the visited point (visual). And iterating until all seed points are accessed, and outputting a set C for finally storing the clustering result if no point which is not accessed exists in the point cloud Q.
In the step S3, in order to prevent the clustering of the objects with too small or too large sizes, the clustering cluster needs to be limited, the method sets a threshold interval of the clustering cluster points, and when the clustering result points are not in the threshold interval, the clusters are not reserved.
In the step S4, by calculating the correlation matrix, the matched dynamic targets in the two-frame point clouds can be determined according to the maximum value of each row of the correlation matrix, however, such matching may cause mismatching between adjacent pedestrians and vehicles, and in order to avoid such problems, some constraints need to be introduced. The most common constraint is that according to the size of the 3D bounding box of two frames of clustering targets, the method suppresses mismatching by judging the relation between the ratio of the volumes of the 3D bounding boxes of the matching targets and a set threshold, and only the targets meeting the formula and the matching result of the incidence matrix simultaneously are considered to be successful in matching, wherein the specific calculation mode is shown as the formula, and l is as follows i ，w i ，h i Representing the length, width and height of the ith object of the point cloud of the previous frame, T v Representing a volume ratio limiting threshold. After the matching of the dynamic targets is completed, the distance between the centers of the matched dynamic targets is calculated, and the dynamic targets in the point cloud of the previous frame are translated to the corresponding positions of the current point cloud according to the distance.
Then experimental results and analysis are carried out
The method adopts a Velodyne 16 line laser radar and a CGI-410 inertial navigation system carried on an unmanned platform independently developed in a laboratory to collect data in a campus and perform continuous frame point cloud space characteristic enhancement detection experiments, the speed of the acquisition is 20km/h, and the acquisition frequency of the laser radar is 10Hz.
Analysis of pretreatment results
Fig. 4 shows a graph of the result of the point cloud preprocessing. The method sets the voxel size of voxel downsampling to be 0.3m, can see the sparse density of the point cloud after preprocessing obviously, removes most points which are not in the ROI area, and is beneficial to subsequent operation. The point cloud before pretreatment contains 24222 points, the point cloud after pretreatment contains 10074 points, and after downsampling and outlier removal, a large number of redundant data points are effectively reduced, and the space geometric information of the point cloud is reserved.
Analysis of ground segmentation results
As shown in fig. 5, which is a graph of ground segmentation results of the point cloud, it can be seen from fig. 4 (a) that the ground information and the space information are well segmented, wherein the red point cloud represents the ground and the yellow point cloud represents the space information. From fig. 4 (b), it can be seen that more complete non-ground information is preserved, and interference of ground factors is filtered for subsequent adaptive clustering. It is worth noting that the number of iterations needs to be carefully chosen, if the number of iterations is too small, the fitted ground plane may lose too much environmental information, and if the number of iterations is too large, may not converge.
Analysis of dynamic target extraction results
Fig. 6 is a graph showing the result of the adaptive clustering of the point clouds, wherein fig. 5 (a) is a result of the adaptive clustering of the point clouds of the previous frame, and fig. 5 (b) is a result of the adaptive clustering of the point clouds of the current frame, wherein red boxes are marked to represent dynamic targets, and the point clouds with different colors represent different categories. According to the method, the clustered point cloud clusters are screened according to actual conditions, most clustering results which do not accord with the actual conditions are deleted, static obstacles are removed according to the fact that the movement distance of clustered targets in the point cloud of the adjacent frame is greater than 0.2m, and the moving objects can be roughly judged to be vehicles and pedestrians.
The continuous frame point cloud space characteristic enhancement result analysis is carried out, and NDT point cloud registration result analysis is carried out first
Fig. 7 is a diagram of NDT point cloud registration results, in which white point cloud represents the previous frame point cloud and yellow point cloud represents the current frame point cloud. The environmental information can be better described by the background information such as tree vegetation in the environment, but the moving vehicles and pedestrians in the environment are greatly displaced and cannot well represent the environmental information, so that the dynamic objects in the environment are required to be treated independently.
Dynamic target point cloud fusion
Fig. 8 is a diagram showing a fusion result of the dynamic target point clouds, wherein the green point cloud represents the dynamic object in the previous frame point cloud, and the yellow point cloud represents the dynamic object in the current frame point cloud. It can be seen from fig. 8 (a) that the dynamic objects in the two frames are significantly displaced if projected directly before the dynamic target point cloud is fused. After nearest neighbor correlation matching, the center of the dynamic target point cloud is moved to the center position of the dynamic target point cloud corresponding to the current frame through translation, and the method is specifically shown in fig. 8 (b). It can be seen that the moved green point cloud and yellow point cloud are not completely overlapped, which means that the thickness of the dynamic target point cloud of the current frame can be effectively increased, and the spatial characteristics of the dynamic target point cloud of the current frame are enhanced.
Static background point cloud fusion
On the basis of completing dynamic target point cloud fusion, projecting an attitude transformation matrix obtained by NDT registration of a background point cloud in a previous frame point cloud to a current frame point cloud, and completing continuous frame point cloud space characteristic enhancement, wherein the specific result is shown in fig. 9. According to the image, the environment point clouds of tree vegetation and the like are basically overlapped, environment information can be reflected well, the dynamic target point cloud also moves to the corresponding position of the current frame point cloud, and the spatial characteristics of the dynamic target are enhanced.
Outputting the detection result
Fig. 10 is a diagram of a detection result of a point cloud after spatial feature enhancement, and the enhanced point cloud is detected by using the adaptive clustering algorithm provided by the method. The method can be used for detecting pedestrians and vehicles on the road well, the density of the point clouds with enhanced spatial characteristics is obviously increased, and the spatial information after two frames of point clouds are fused is basically overlapped. The pedestrian with sparse edges can be better detected by using the fusion point cloud for target detection, which indicates that the fusion point cloud has better effect on a single target.
According to the performance analysis, real-time tests are carried out by utilizing 373 frames of point cloud and inertial navigation data of 38.5s acquired by velodyne VLP-16 laser radar and CGI-410 inertial navigation, and the detection performance of the point cloud after NDT point cloud registration and fusion is analyzed. As shown in fig. 3, which is a NDT point cloud registration performance chart, it can be seen from the chart that the fitness (fitness) and rmse (root mean square error) of the point cloud registration can more stably float between a certain threshold, and the fitness overall tends to increase, and the rmse average value is 0.16707. The closer the fitness is to 1, the closer the rmse is to 0, which shows that the better the registration effect is, the fluctuation of the registration effect shows that the IMU data acquired by inertial navigation in the registration process is unstable, and the registration effect can meet the requirements of a continuous frame point cloud space feature enhancement algorithm in general.
The method also respectively utilizes an object detection algorithm based on European clustering and an object detection algorithm based on continuous frame space characteristic enhancement to carry out experiments on the acquired data, and the performance pairs are shown in a table 2. It can be seen from the table that the detection speed of the algorithm based on the euclidean clustering is faster, because the algorithm based on the continuous frame space feature enhancement needs to estimate the initial pose transformation matrix and the NDT registration, which takes a long time. But the detection accuracy based on the enhancement of the spatial features of the continuous frames is 11.7% higher than that of the algorithm based on the enhancement and the algorithm based on the euclidean clustering.
Table 2 comparison table of detection algorithm performance
According to the experimental results, the spatial density of a target object is effectively improved, the spatial characteristics of point cloud are enhanced, the accuracy of target detection is improved, the robustness of inertial navigation measurement data is increased, the method has important application value to the unmanned perception field, and aiming at the problems of low offset error and robustness and the like of an IMU estimated posture transformation matrix, the method is used for carrying out fine registration on the IMU estimated posture transformation by using an NDT (Normal Distribution Transformer, NDT) registration algorithm, so that the error caused by integral accumulation of the IMU is reduced. And then extracting a dynamic target by using a traditional target detection algorithm, independently processing and fusing, and projecting the background point cloud of the previous frame to the point cloud of the current frame according to the estimated gesture transformation matrix, so that the density of the sparse point cloud is improved, and the accuracy of obstacle detection is improved.
It should be noted that in the present method, relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.
The above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims (8)
1. An obstacle detection method based on continuous frame point cloud space feature enhancement is characterized by comprising the following steps of: the detection method based on continuous frame point cloud space feature enhancement comprises the following steps: s1: firstly, performing point cloud pretreatment, wherein the point cloud pretreatment mainly comprises three parts of downsampling, outlier removal and ROI (Region of Interest) area extraction; in order to accelerate the operation speed and reduce the storage consumption, the point cloud of the dense area needs to be downsampled, and the characteristic space is ensured to be unchanged;
s2: ground segmentation is carried out, ground information and space information are separated, and subsequent point cloud clustering and obstacle recognition are facilitated;
s3: extracting a dynamic target, namely extracting the dynamic target by using a DBSCAN (Density-Based Clustering Algorithm with Noise, DBSCAN) clustering algorithm;
s4: performing dynamic target point cloud fusion, and after extracting dynamic targets of two adjacent frames of point clouds, performing data association matching on the dynamic targets in the two frames of point clouds; the method adopts a nearest neighbor data association (Nearest Neighbor Data Association, NNDA) method to realize the matching between moving objects between two frames of point clouds; the method comprises calculating an incidence matrix according to Euclidean distance between obstacles, taking maximum value of each row in the incidence matrix as incidence result of the obstacles, assuming N objects exist in clustering result of the previous frame point cloud, and M objects exist in clustering result of the current frame point cloud, the incidence matrix can be expressed as formula, wherein D is as follows i,j Representing the Euclidean distance between the ith object in the point cloud of the previous frame and the jth object in the point cloud of the current frame;
s5, carrying out static background point cloud fusion, using an inertial navigation system of CGI-410 to obtain triaxial acceleration and attitude angle of the vehicle, adopting an NDT algorithm as an attitude transformation matrix between continuous frame point clouds, and using the triaxial acceleration and attitude angle provided by the inertial navigation system to estimate an initial transformation matrix, thereby eliminating the problem of weak robustness in the registration process; assume that the inertial navigation system is at t 1 The three-axis acceleration obtained at the moment is a 1 Three-axis speed v 1 Attitude angle θ 1 At t 2 The acceleration obtained at the moment is a 2 Attitude angle θ 2 Wherein each variable contains components of the x, y, z axes; the acceleration increased in the process is delta a, the changed attitude angle is delta theta, a rotation matrix R of the vehicle attitude change at two moments can be calculated according to a formula, a translation vector t is calculated through the formula, and the finally solved initial transformation matrix init_trans is shown in the formula;
s6, firstly rasterizing the point clouds by using an NDT algorithm, and representing the point cloud distribution in each grid by using a normal distribution probability function, wherein the normal distribution probability density function is calculated in a mode shown in the formula, and the normal distribution probability function is calculated by using a standard distribution probability functionRepresenting a normal distribution probability density function,represents the mean vector, Σ represents the covariance matrix, +.>Representing a point cloud within a grid, c 1 Is a constant;
after the probability density function is obtained, mapping the point cloud to the target point cloud through initialized transformation parameters, determining normal distribution corresponding to each mapping point, and summing the probability density of the mapping points to obtain a score value s (p) of each coordinate transformation parameter, wherein the score value s (p) is shown as a formulaRepresenting a point cloud vector within the mapped grid, +.>Representing the mapped mean vector; then, a Hessian matrix method is utilized to calculate the minimum value of s (p), and finally, an Newton iteration method is utilized until an optimization function converges to calculate the optimal estimation of the transformation parameters;
after the optimal transformation matrix is obtained through the NDT, the point cloud of the previous frame can be projected to the point cloud of the current frame through rotation and translation, so that fusion of the background point cloud is realized.
2. The detection method based on continuous frame point cloud space feature enhancement as claimed in claim 1, wherein: in the step S, the method can rapidly and effectively pump the point cloudThin, is widely applied to three-dimensional point cloud processing; the method comprises the steps of firstly establishing an axial bounding box for a three-dimensional point cloud, dividing the bounding box into n equal parts along each coordinate axis direction to obtain n multiplied by n voxels, selecting the mass center of the voxel from each voxel to represent the voxel, and completing voxel grid downsampling, wherein the mass center selecting method is shown as the formula, wherein (x c ,y c ,z c ) Represents centroid coordinates, n represents the total number of voxel midpoints, (x) i ,y i ,z i ) Representing coordinates of an i-th point in the voxel; the voxel grid method downsampling has the advantages of high calculation efficiency, uniform distribution of sampling points, effective reduction of data volume and acceleration of operation speed;
3. the detection method based on continuous frame point cloud space feature enhancement as claimed in claim 1, wherein: in the step S1, filtering of the point cloud outliers is implemented by using a point cloud radius filtering method; the basic idea of the method is that each data point in the initial point cloud is assumed to have at least a specified number of neighborhood points in a specified radius neighborhood, the points meeting the conditions are regarded as normal point reservation, and the rest points are regarded as outlier removal; for the ith point p in the point cloud containing N points i Calculated as p i The sphere center, r is the number n of points in the sphere with radius, if the number of points in the neighborhood is smaller than a certain threshold T r If the point is determined to be an outlier, otherwise, the point is regarded as a normal point to be reserved, and the specific calculation is shown in the formula; the method sets T r 20, storing point cloud data by using a KD Tree (K-dimension Tree) data structure so as to realize rapid search of neighborhood points;
according to the characteristic of near-dense and far-sparse of the collected point cloud, the method sets the extracted point cloud ROI area to be [ -70m,70m ] along the x-axis, to be [ -20m,20m ] along the y-axis, and limits the height along the z-axis to be 2.5m; and extracting indexes of points in the ROI by using the established KD tree data structure, so that the segmentation of the ROI is completed, and the speed of point cloud segmentation is increased.
4. The detection method based on continuous frame point cloud space feature enhancement as claimed in claim 1, wherein: in the step S2, since the real ground often has many slopes, which is not a perfect plane, the ground needs to be processed when the ground with a certain slope change is needed to be well segmented; the method utilizes an algorithm based on a plane model to fit the ground, the idea of the algorithm is that firstly, a space is divided into N sub-planes along the x-axis direction (the head direction) of a point cloud, the points in the ith sub-plane are firstly ordered according to the size of a z-axis, l points in the ith sub-plane are selected to form a seed point set, and the plane normal vector of the ith sub-plane is estimated through singular value decompositionThereby obtaining the sub-plane fitting ground sigma i The method comprises the steps of carrying out a first treatment on the surface of the Then, carrying out orthogonal projection judgment on each point in the ith sub-plane by using an iterative method, wherein the specific method is that if the point reaches the estimated ground sigma i Is smaller than threshold th d Then consider the point to be a ground point and add the point to sigma i Otherwise, do not join, thereby iterating N iter Fitting ground sigma of ith sub-plane can be obtained i Finally, the ground fitted by each sub-plane is spliced together to form a final estimated ground plane sigma, wherein the fitted plane can be represented as shown in a formula, d represents an intercept, and the final estimated ground can be represented as shown in the formula;
σ＝[σ 1 ,L,σ N ] (4)。
5. the detection method based on continuous frame point cloud space feature enhancement as claimed in claim 1, wherein: in the step S3, in order to accelerate the clustering process, the point cloud is often projected to a horizontal plane to perform clustering, so as to reduce the operand in the clustering process; because the DBSCAN algorithm is sensitive to the neighborhood radius and the minimum point threshold, the method firstly divides the point cloud into M sub-point clouds according to the characteristics of near-dense and far-sparse point clouds, constructs KD tree data structures in each sub-point cloud to improve the efficiency of a clustering algorithm, and sets different clustering radiuses and the minimum point threshold according to the difference of the distances from the center of each sub-point cloud to the origin of the laser radar so as to realize self-adaptive DBSCAN clustering, wherein the clustering radiuses and the distances from the sub-point clouds to the origin of the laser radar form a negative correlation relation, and a specific relational expression is shown in a formula (x) j ,x j ) Represents the jth point, epsilon, in the ith sub-point cloud i Representing the clustering neighborhood radius of the ith sub-point cloud, S i Representing an ith sub-point cloud;
6. the detection method based on continuous frame point cloud space feature enhancement as claimed in claim 1, wherein: in the step S3, for the input segmented sub-point cloud, all points in the point cloud are marked as unvisited points (unvisited), and then a storage structure Q of the point cloud is constructed by using a KD tree, so as to facilitate subsequent operations; randomly taking a point P from the point cloud, and judging whether the point P is a core point or not by calculating the distance between the point P and a neighborhood point; if the point is a core point, the point can be mutually communicated with other core points, a new cluster is created, the point is added into the cluster, all points in the neighborhood of the point are taken as a seed point set, the core points are searched in the seed point set in an iterative mode, namely, whether the point is the core point or not is judged by randomly taking one point in the seed point set, if yes, the neighborhood point of the point is taken as the seed point, and if not, the point is marked as the visited point (visited); and iterating until all seed points are accessed, and outputting a set C for finally storing the clustering result if no point which is not accessed exists in the point cloud Q.
7. The detection method based on continuous frame point cloud space feature enhancement as claimed in claim 1, wherein: in the step S3, in order to prevent the clustering of the objects with too small or too large sizes, the clustering cluster needs to be limited, the method sets a threshold interval of the clustering cluster points, and when the clustering result points are not in the threshold interval, the clusters are not reserved.
8. The detection method based on continuous frame point cloud space feature enhancement as claimed in claim 1, wherein: in the step S4, by calculating the correlation matrix, the matched dynamic targets in the two-frame point clouds can be determined according to the maximum value of each row of the correlation matrix, however, such matching may cause mismatching between adjacent pedestrians and vehicles, and in order to avoid such problems, some constraint conditions need to be introduced; the most common constraint is that according to the size of the 3D bounding box of two frames of clustering targets, the method suppresses mismatching by judging the relation between the ratio of the volumes of the 3D bounding boxes of the matching targets and a set threshold, and only the targets meeting the formula and the matching result of the incidence matrix simultaneously are considered to be successful in matching, wherein the specific calculation mode is shown as the formula, and l is as follows i ，w i ，h i Representing the length, width and height of the ith object of the point cloud of the previous frame, T v Representing a volume ratio limit threshold; after the matching of the dynamic targets is completed, the distance between the centers of the matched dynamic targets is required to be calculated, and the dynamic targets in the previous frame of point cloud are translated to the corresponding positions of the current point cloud according to the distance;
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311703585.9A CN117830932A (en) | 2023-12-12 | 2023-12-12 | Obstacle detection method based on continuous frame point cloud space feature enhancement |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311703585.9A CN117830932A (en) | 2023-12-12 | 2023-12-12 | Obstacle detection method based on continuous frame point cloud space feature enhancement |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117830932A true CN117830932A (en) | 2024-04-05 |
Family
ID=90521893
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311703585.9A Pending CN117830932A (en) | 2023-12-12 | 2023-12-12 | Obstacle detection method based on continuous frame point cloud space feature enhancement |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117830932A (en) |
-
2023
- 2023-12-12 CN CN202311703585.9A patent/CN117830932A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111583369B (en) | Laser SLAM method based on facial line angular point feature extraction | |
| CN112101092A (en) | Automatic driving environment sensing method and system | |
| CN113156421A (en) | Obstacle detection method based on information fusion of millimeter wave radar and camera | |
| CN113345008B (en) | Laser radar dynamic obstacle detection method considering wheel type robot position and posture estimation | |
| CN113506318B (en) | Three-dimensional target perception method under vehicle-mounted edge scene | |
| CN113345018A (en) | Laser monocular vision fusion positioning mapping method in dynamic scene | |
| CN110674705A (en) | Small-sized obstacle detection method and device based on multi-line laser radar | |
| Wang et al. | An overview of 3d object detection | |
| CN112578673B (en) | Perception decision and tracking control method for multi-sensor fusion of formula-free racing car | |
| CN114359876B (en) | Vehicle target identification method and storage medium | |
| CN116109601A (en) | Real-time target detection method based on three-dimensional laser radar point cloud | |
| Wen et al. | Research on 3D point cloud de-distortion algorithm and its application on Euclidean clustering | |
| CN114325634A (en) | Method for extracting passable area in high-robustness field environment based on laser radar | |
| CN115620261A (en) | Vehicle environment sensing method, system, equipment and medium based on multiple sensors | |
| CN115861968A (en) | Dynamic obstacle removing method based on real-time point cloud data | |
| Qing et al. | A novel particle filter implementation for a multiple-vehicle detection and tracking system using tail light segmentation | |
| Zhou et al. | A lidar mapping system for robot navigation in dynamic environments | |
| CN114066773B (en) | Dynamic object removal based on point cloud characteristics and Monte Carlo expansion method | |
| CN114241448A (en) | Method and device for obtaining heading angle of obstacle, electronic equipment and vehicle | |
| Liu et al. | A lightweight lidar-camera sensing method of obstacles detection and classification for autonomous rail rapid transit | |
| CN117576665B (en) | Automatic driving-oriented single-camera three-dimensional target detection method and system | |
| CN113536959A (en) | Dynamic obstacle detection method based on stereoscopic vision | |
| CN118411507A (en) | Semantic map construction method and system for scene with dynamic target | |
| Zhang et al. | Research on mobile robot target recognition and obstacle avoidance based on vision | |
| CN117269952A (en) | Method and device for semi-automatically labeling moving target point cloud of 4D imaging millimeter wave radar |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |