CN108596053A - A kind of vehicle checking method and system based on SSD and vehicle attitude classification - Google Patents
A kind of vehicle checking method and system based on SSD and vehicle attitude classification Download PDFInfo
- Publication number
- CN108596053A CN108596053A CN201810311110.8A CN201810311110A CN108596053A CN 108596053 A CN108596053 A CN 108596053A CN 201810311110 A CN201810311110 A CN 201810311110A CN 108596053 A CN108596053 A CN 108596053A
- Authority
- CN
- China
- Prior art keywords
- vehicle
- loss
- detection
- task
- target frame
- 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.)
- Granted
Links
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
- G06V20/54—Surveillance or monitoring of activities, e.g. for recognising suspicious objects of traffic, e.g. cars on the road, trains or boats
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/017—Detecting movement of traffic to be counted or controlled identifying vehicles
- G08G1/0175—Detecting movement of traffic to be counted or controlled identifying vehicles by photographing vehicles, e.g. when violating traffic rules
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of vehicle checking method classified based on SSD and vehicle attitude and system, the method includes：Vehicle attitude is divided according to the angle of headstock and trunnion axis, vehicle attitude classification task is added on original SSD network models, vehicle detection loss and the loss of vehicle attitude classification task are combined to form multitask loss, the softmax losses of original SSD models are replaced with into focal loss losses, vehicle attitude classification task and vehicle detection task cooperative are optimized, training obtains detection model, vehicle detection is carried out to picture to be detected using detection model, realizes multiple dimensioned, multi-angle vehicle detection.Deep learning target detection SSD is used for vehicle detection by the present invention, it is trained using vehicle attitude classification as nonproductive task and vehicle detection task cooperative, and add focal loss and solve the problems, such as vehicle sample imbalance, to improve the Stability and veracity of system.
Description
Technical field
The invention belongs to image procossings and pattern classification field, more particularly, to one kind based on SSD and vehicle attitude point
The vehicle checking method and system of class.
Background technology
With the rapid development of artificial intelligence and computer vision, intelligent transportation system becomes the hair of Modern Traffic system
Direction is opened up, wherein vehicle detection is the important component of intelligent transportation system.Vehicle detection is broadly divided into two large divisions.First
Part is that the vehicle based on video detects in real time, and the moving object detection based on video is that one of computer vision research is important
Research direction, it is the basis further studied that moving vehicle target is fast and accurately separated from image, further
It is the tracking to moving vehicle, many traffic informations can be obtained, has in the systems such as traffic monitoring, traffic incidents detection wide
General application；Second part is the vehicle detection based on single picture, due to not having video sequence, so vehicle can not be obtained
Movable information can only detect vehicle from the visual information of static images.
Vehicle detection based on video, common thinking is that prominent vehicle target either eliminates background, common to move
Vehicle checking method mainly has：Frame differential method, background subtraction, optical flow etc..Vehicle detection based on single image passes
The method of system is mainly based upon vehicle edge detection, when that cannot extract ideal vehicle edge, such as vehicle color and road face
Color is similar or when congested with cars occurs.In recent years, with the prevalence of depth convolutional neural networks, many is based on depth
The object detection method of convolutional network all achieves surprising accuracy of detection.Current state-of-the-art object detecting method is broadly divided into
Two classes, one kind are to be based on extracted region, such as R-CNN, Fast R-CNN and Faster R-CNN etc..These methods are to be divided to two ranks
Section, so time performance is poor, it is difficult to real time execution.
Invention content
For the disadvantages described above or Improvement requirement of the prior art, the present invention provides one kind based on SSD and vehicle attitude point
Thus the vehicle checking method and system of class solve the technical problems such as the low, poor robustness of accuracy rate present in existing method.
To achieve the above object, according to one aspect of the present invention, it provides and a kind of is classified based on SSD and vehicle attitude
Vehicle checking method, including：
(1) classified to vehicle attitude according to the angle of headstock and trunnion axis, and be added on original SSD network models
Vehicle attitude classification task, wherein the vehicle attitude classification task and vehicle detection task are arranged side by side；
(2) it after default weight coefficient being multiplied by the loss of the vehicle attitude classification task of addition, is examined with the vehicle
The loss of survey task is added to form multitask loss；
(3) loss of the focal loss loss as the vehicle attitude classification task of addition is used, and by focal
Loss loses the Classification Loss as the vehicle target in the loss of the vehicle detection task；
(4) the vehicle detection network based on SSD that the vehicle attitude classification task is added is built, is examined according to the vehicle
It includes vehicle that survey grid network, which is generated on the characteristic pattern of different layers in the candidate target frame and each candidate target frame of fixed size,
Possibility, sorted to all candidate target frame according to the confidence level that each candidate target frame includes vehicle, and remove and set
The low negative sample of reliability calculates described more so that the ratio of positive negative sample is maintained at preset ratio according to the positive negative sample of reservation
Task is lost, then backpropagation, repetitive exercise, until reaching maximum iteration obtains target detection model, with by described
Target detection model carries out vehicle detection to picture to be detected.
In the present invention, vehicle attitude is divided by front, the back side and side three classes using the angle of headstock and trunnion axis, made
Obtaining vehicle attitude can quantify, and then classify.
In the present invention, nonproductive task of the vehicle attitude classification task as vehicle detection task is added, correlation is passed through
The principle that can be mutually promoted between task effectively increases vehicle detection precision.
In the present invention, a weight coefficient is multiplied by the loss of the vehicle attitude classification task of addition then to examine with vehicle
The loss of survey task is added, and weight coefficient here has selected a suitable value by cross-validation experiments, to effectively carry
High system accuracy.
In the present invention, by the loss of the Classification Loss of original vehicle target and the posture classification task being newly added all by
Common softmax losses replace with focal loss losses, solve the problems, such as class imbalance, to effectively increase system
Precision.
Preferably, the multitask, which is lost, is：
Wherein, Lconf(x, c) is indicated
The target frame of prediction belongs to vehicle or belongs to the loss of background, Lloc(x, l, g) indicates what the coordinate of the target frame of prediction returned
Loss, α account for the weight coefficient of total losses, L for itpose(x,cp) indicating the loss of vehicle attitude classification task, β accounts for total damage for it
The weight coefficient of mistake, x indicate whether the target frame of prediction can match with the true value frame of label, are 1 if successful match, matching
Failure then indicates that target frame belongs to the probability of vehicle for 0, c, and l indicates that the vehicle target frame of prediction, g indicate true value frame, N expressions
The number for the default frame matched, cpIndicate that target frame belongs to the probability of some vehicle attitude classification.
Preferably, step (4) includes：
(4.1) the vehicle detection network based on SSD that vehicle attitude classification task is added is built, wherein the vehicle inspection
Survey grid network uses VGGNet, after pre-training is complete on ImageNet data sets, is replaced subsequent two layers with two new convolutional layers
Full articulamentum, and increase by 4 additional convolutional layers；
(4.2) using several training pictures as input, it is based on the vehicle detection network, on the characteristic pattern of different layers
Generate the possibility for including vehicle in the candidate target frame and each candidate target frame of several fixed sizes；
(4.3) include that vehicle confidence level sorts to all candidate target frames, and removes and set according to each candidate target frame
The low negative sample of reliability, so that the ratio of positive negative sample is maintained at preset ratio；
(4.4) the recurrence loss of the Classification Loss, vehicle target frame coordinate of the vehicle target of the positive negative sample retained is calculated
And the loss of vehicle attitude classification task, and the recurrence of the Classification Loss of the vehicle target, vehicle target frame coordinate is damaged
Summation obtains multitask loss after respective default weight is multiplied by mistake and the loss of vehicle attitude classification task；
(4.5) backpropagation is lost to the multitask, updates network parameter, and repeat step (4.2)~step
(4.4), until network iterations reach default iterations, target vehicle detection model is obtained.
Preferably, described to include to picture to be detected progress vehicle detection by the target detection model：
Picture to be detected is handled by the target detection model, to generate the candidate target of several fixed sizes
Include the confidence level of vehicle in frame and each candidate target frame；
Non-maxima suppression operation is carried out to candidate target frame, removes the candidate target frame of repetition, it is final to obtain
Testing result.
It is another aspect of this invention to provide that providing a kind of vehicle detecting system classified based on SSD and vehicle attitude, packet
It includes：
Vehicle attitude sort module, for being classified to vehicle attitude according to the angle of headstock and trunnion axis, and in original
Vehicle attitude classification task is added on beginning SSD network model, wherein the vehicle attitude classification task and vehicle detection task are simultaneously
Row；
Multitask loss structure module, for default weight to be multiplied by the loss for the vehicle attitude classification task being added
After coefficient, it is added to form multitask loss with the loss of the vehicle detection task；
Loss function determining module, for the vehicle attitude classification task using focal loss losses as addition
Loss, and focal loss are lost into the Classification Loss as the vehicle target in the loss of the vehicle detection task；
Joint training module, for building the vehicle detection network based on SSD that the vehicle attitude classification task is added,
Candidate target frame and each candidate of fixed size are generated on the characteristic pattern of different layers according to the vehicle detection network
The possibility for including vehicle in target frame, according to confidence level of each candidate target frame comprising vehicle to all candidate target frames
Sequence, and the low negative sample of confidence level is removed, so that the ratio of positive negative sample is maintained at preset ratio, according to the positive and negative sample of reservation
This calculating multitask loss, then backpropagation, repetitive exercise, until reaching maximum iteration obtains target detection mould
Type, to carry out vehicle detection to picture to be detected by the target detection model.
In general, through the invention it is contemplated above technical scheme is compared with the prior art, can obtain down and show
Beneficial effect：
(1) as a result of SSD algorithms so that algorithm, can be with real time execution while precision is very high；
(2) due to assisting vehicle detection using vehicle attitude classification task, accuracy of detection is effectively increased；
(3) due to adding focal loss, efficiently solve the problems, such as between different posture samples quantity it is unbalanced with
And the unbalanced problem of difficulty or ease sample size, so that accuracy of detection is further enhanced.
Description of the drawings
Fig. 1 is a kind of flow of vehicle checking method classified based on SSD and vehicle attitude provided in an embodiment of the present invention
Schematic diagram；
Fig. 2 is a kind of network of vehicle detecting system classified based on SSD and vehicle attitude provided in an embodiment of the present invention
Structural schematic diagram；
Fig. 3 is that a kind of vehicle attitude provided in an embodiment of the present invention divides schematic diagram.
Specific implementation mode
In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.As long as in addition, technical characteristic involved in the various embodiments of the present invention described below
It does not constitute a conflict with each other and can be combined with each other.
The invention discloses a kind of vehicle checking method classified based on SSD and vehicle attitude and systems, are suitable for picture
It is middle there are when multiple dimensioned, multi-angle vehicle, can accurately detect vehicle location.It is examined more particularly to based on deep learning target
The vehicle detection of method of determining and calculating SSD and vehicle attitude classification, it is main to use object detection method, vehicle detection and vehicle based on SSD
The multitask training method that the classification of posture combines and the method for solving class imbalance using focal loss, in picture
Vehicle detection, blocked in various degree in particular for having more angles, scale different in same pictures and existing
The case where when vehicle.
The term used first to the present invention below is explained and illustrated.
SSD：SSD is the abbreviation of Single Shot MultiBox Detector.SSD algorithms are a kind of directly prediction mesh
The algorithm of target detection for marking the coordinate and classification of frame, does not generate the process of candidate frame.The master network structure of algorithm is VGG16,
Make two full articulamentums into convolutional layer and is further added by 4 convolutional layer tectonic network structures.
Bounding box are returned：It is to carry out finer adjustment to the target window detected that Bounding box, which are returned,
Allow it to the physical location for being more nearly target.Window is generally used four dimensional vectors (center point coordinate of window and
It is wide high) it indicates, Bounding box, which are returned, first to carry out translation transformation to the center point coordinate of window, then high to the width of window
Carry out scaling so that window is closer to target to be detected.
Softmax：Softmax is used for different object allocation probability values.Its input is a vector, and output is also one
A vector, each of input vector value is treated as power exponent evaluation, then regularization these end values by it, you can output vector.
As shown in Figure 1, for the flow diagram for the vehicle checking method classified the present invention is based on SSD and vehicle attitude, this
The method of invention specifically includes following steps：
(1) prepare data set, data set includes the image containing vehicle and corresponding mark file, wherein marking file
The information of mark is other than usually detecting required target category and target frame information, it is also necessary to the headstock of vehicle target with
The angle of trunnion axis.
(2) vehicle attitude is divided, vehicle attitude is divided by front, the back side and side according to the angle of headstock and trunnion axis
Three classes.
(3) network model is defined, vehicle attitude classification task is added on original SSD network models.Vehicle attitude is vehicle
Important attribute, the vehicle attitude classification task and vehicle detection task of addition be arranged side by side, wherein vehicle detection task includes vehicle
The classification task of target and the recurrence task of vehicle target frame, they share the convolutional layer of front in original SSD network models.
(4) multitask loss is defined, by the vehicle attitude classification task of loss and the addition of original vehicle detection task
Loss combine to form multitask loss, the wherein loss of vehicle attitude classification is also the softmax losses of multiclass.
Specifically combination is：The loss of original vehicle Detection task includes the Classification Loss of vehicle target and returning for vehicle target frame
Return loss, then a weight coefficient is multiplied by the loss of the vehicle attitude classification task of addition to be added with the loss of front two.
The loss of original vehicle detection is as follows：
It is as follows that the sorted loss of vehicle attitude is added：
Wherein, Lconf(x, c) indicates that the target frame of prediction belongs to vehicle or belongs to the loss of background；Lloc(x, l, g) table
Show that the loss that the coordinate of the target frame of prediction returns, α account for the weight coefficient of total losses for it；Lpose(x,cp) indicate vehicle attitude
The loss of classification task, β account for the weight coefficient of total losses for it.X indicates whether the target frame of prediction can be with the true value frame of label
Matching is 1 if successful match, indicates that target frame belongs to the probability of vehicle if it fails to match for 0, c, l indicates the vehicle of prediction
Target frame, g indicate that true value frame, N indicate the number of matched default frame, cpIndicate that target frame belongs to some vehicle attitude class
Other probability.
Wherein parameter alpha and β can be determined by cross-validation experiments.
(5) Classification Loss function is replaced.Since there are unbalanced problem, institutes for the number of vehicles of different posture in data set
With the loss for the vehicle attitude classification task being newly added, we can be used for solving class using focal loss, focal loss
Other imbalance problem.Meanwhile the Classification Loss of the vehicle target in the loss of former vehicle detection task loses for softmax, and
Vehicle target has that difficulty divides sample and easily divides sample imbalance, we also replace with original softmax losses
Focal loss losses, to solve the problems, such as difficulty or ease sample imbalance.
(6) joint training optimizes vehicle attitude classification task and vehicle detection task cooperative.According to the vehicle put up
Detection network generates the candidate target frame and each candidate target frame of a series of fixed sizes on the characteristic pattern of different layers
In include the possibility of vehicle.Then all candidate target frames are arranged according to confidence level of each candidate target frame comprising vehicle
Sequence, and the low negative sample of confidence level is removed, so that the ratio of positive negative sample is maintained at preset ratio.It, will be new when counting loss
The loss of the vehicle attitude classification task of addition be multiplied by a weight coefficient again with the vehicle target of former vehicle detection task point
Class is lost and the recurrence of vehicle target frame loss summation, then backpropagation, repetitive exercise, until reaching greatest iteration time
Number.Specifically include following sub-step：
(6-1) builds the vehicle detection network based on SSD that vehicle attitude classification task is added：Network uses VGGNet,
After pre-training is complete on ImageNet data sets, subsequent two layers full articulamentum is replaced with two new convolutional layers, and increase by 4
Additional convolutional layer.
(6-2) inputs a certain amount of trained picture, based on the vehicle detection network in step (6-1), in the spy of different layers
A series of possibility for including vehicle in the candidate target frame and each candidate target frame of fixed sizes is generated on sign figure.
(6-3) results in positive and negative sample since the most candidate target frames extracted in step (6-2) are all negative samples
This imbalance.It is sorted to all candidate target frames according to the confidence level that each candidate target frame includes vehicle, and removes confidence
Low negative sample is spent, the ratio of positive negative sample is finally made to be maintained at preset ratio.
In embodiments of the present invention, Hard Negative Mining strategies may be used, according to each candidate target frame
Including the confidence level of vehicle sorts to all candidate target frames, and the low negative sample of confidence level is removed, finally makes positive negative sample
Ratio be maintained at preset ratio.
Wherein, preset ratio can be determined according to actual needs, and in embodiments of the present invention, preset ratio is preferably
1:3。
(6-4) calculates Classification Loss, the vehicle target frame coordinate of the vehicle target of the positive negative sample retained in step (6-3)
Return loss and the loss of vehicle attitude classification task.The Classification Loss of vehicle target indicate candidate target frame be vehicle or
It is not the loss of vehicle；It is the loss finely returned to the coordinate of candidate target frame that vehicle target frame coordinate, which returns loss,；
The loss of vehicle attitude classification task is that the loss of vehicle attitude classification is carried out to target frame.The wherein Classification Loss of vehicle target
Loss with vehicle attitude classification task uses focal loss forms, solves the problems, such as sample imbalance.Then three are lost
It sums after being multiplied by respective weight.
The loss backpropagation of (6-5) to being calculated in step (6-4) updates network parameter.
(6-6) repeats step (6-2) and arrives (6-4), until network iterations reach default iterations to get to target
Vehicle detection model.
Wherein, default iterations can be determined according to actual conditions.
(7) vehicle detection is carried out to picture to be detected using the target vehicle detection model that training obtains in step (6), from
And realize multiple dimensioned, multi-angle vehicle detection.Specifically include following sub-step：
(7-1) is loaded into the target vehicle detection model that training obtains in step (6).
(7-2) inputs picture to be detected.
The target vehicle detection model that (7-3) is based in step (7-1) generates the candidate target frame of several fixed sizes,
And include the confidence level of object example in each target frame.
(7-4) carries out non-maxima suppression operation to the candidate target frame in step (7-3), removes the target candidate of repetition
Frame, to obtain final testing result.
Fig. 2 is the network structure for the vehicle detecting system classified the present invention is based on SSD and vehicle attitude, is wrapped in fig. 2
Include input picture module, characteristic extracting module and classification and regression block.Wherein input picture module is by each mapping to be checked
Fixed size as all zooming to 300*300；Characteristic extracting module carries out multilayer process of convolution to the image of input, to obtain
Multiple characteristic patterns of different scale and size；Classification and regression block choose several specific characteristic patterns, then using fixed big
Small convolution kernel goes the classification and coordinate shift information of prediction vehicle target.
Fig. 3 is that the vehicle attitude for the vehicle detecting system classified the present invention is based on SSD and vehicle attitude divides schematic diagram.I
Vehicle attitude is divided by front, the back side and side three classes according to the angle of headstock and trunnion axis, Fig. 3 is respectively front, the back of the body
The vehicle sample schematic diagram in face and side.
The present invention provides a kind of vehicle checking method classified based on SSD and vehicle attitude and systems, preferably solve
Multiple dimensioned, multi-angle and there is the vehicle detection situation mutually blocked, using vehicle attitude classification as nonproductive task and
Vehicle detection task cooperative is trained, and is added focal loss and solved the problems, such as vehicle sample imbalance, to improve system
Stability and veracity.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, all within the spirits and principles of the present invention made by all any modification, equivalent and improvement etc., should all include
Within protection scope of the present invention.
Claims (5)
1. a kind of vehicle checking method classified based on SSD and vehicle attitude, which is characterized in that including：
(1) classified to vehicle attitude according to the angle of headstock and trunnion axis, and vehicle is added on original SSD network models
Posture classification task, wherein the vehicle attitude classification task and vehicle detection task are arranged side by side；
(2) after default weight coefficient being multiplied by the loss of the vehicle attitude classification task of addition, appoint with the vehicle detection
The loss of business is added to form multitask loss；
(3) loss of the focal loss loss as the vehicle attitude classification task of addition is used, and by focal loss
Lose the Classification Loss as the vehicle target in the loss of the vehicle detection task；
(4) the vehicle detection network based on SSD that the vehicle attitude classification task is added is built, according to the vehicle detection net
What network was generated in the candidate target frame and each candidate target frame of fixed size on the characteristic pattern of different layers comprising vehicle
Possibility sorts to all candidate target frames according to the confidence level that each candidate target frame includes vehicle, and removes confidence level
Low negative sample calculates the multitask so that the ratio of positive negative sample is maintained at preset ratio according to the positive negative sample of reservation
Loss, then backpropagation, repetitive exercise, until reaching maximum iteration obtains target detection model, with by the target
Detection model carries out vehicle detection to picture to be detected.
2. according to the method described in claim 1, it is characterized in that, multitask loss is：
Wherein, Lconf(x, c) is indicated
The target frame of prediction belongs to vehicle or belongs to the loss of background, Lloc(x, l, g) indicates what the coordinate of the target frame of prediction returned
Loss, α account for the weight coefficient of total losses, L for itpose(x,cp) indicating the loss of vehicle attitude classification task, β accounts for total damage for it
The weight coefficient of mistake, x indicate whether the target frame of prediction can match with the true value frame of label, are 1 if successful match, matching
Failure then indicates that target frame belongs to the probability of vehicle for 0, c, and l indicates that the vehicle target frame of prediction, g indicate true value frame, N expressions
The number for the default frame matched, cpIndicate that target frame belongs to the probability of some vehicle attitude classification.
3. method according to claim 1 or 2, which is characterized in that step (4) includes：
(4.1) the vehicle detection network based on SSD that vehicle attitude classification task is added is built, wherein the vehicle detection net
Network uses VGGNet, after pre-training is complete on ImageNet data sets, replaces subsequent two layers to connect entirely with two new convolutional layers
Layer is connect, and increases by 4 additional convolutional layers；
(4.2) using several training pictures as input, it is based on the vehicle detection network, is generated on the characteristic pattern of different layers
Include the possibility of vehicle in the candidate target frame of several fixed sizes and each candidate target frame；
(4.3) include that vehicle confidence level sorts to all candidate target frames, and removes confidence level according to each candidate target frame
Low negative sample, so that the ratio of positive negative sample is maintained at preset ratio；
(4.4) calculate the Classification Loss of vehicle target, vehicle target frame coordinate of the positive negative sample retained recurrence loss and
The loss of vehicle attitude classification task, and by the recurrence of the Classification Loss of the vehicle target, vehicle target frame coordinate lose with
And the loss of vehicle attitude classification task is multiplied by summation after respective default weight and obtains multitask loss；
(4.5) backpropagation is lost to the multitask, updates network parameter, and repeat step (4.2)~step
(4.4), until network iterations reach default iterations, target vehicle detection model is obtained.
4. according to the method described in claim 3, it is characterized in that, it is described by the target detection model to picture to be detected into
Row vehicle detection includes：
Picture to be detected is handled by the target detection model, to generate the candidate target frame of several fixed sizes,
And include the confidence level of vehicle in each candidate target frame；
Non-maxima suppression operation is carried out to candidate target frame, removes the candidate target frame of repetition, to obtain final detection
As a result.
5. a kind of vehicle detecting system classified based on SSD and vehicle attitude, which is characterized in that including：
Vehicle attitude sort module, for being classified to vehicle attitude according to the angle of headstock and trunnion axis, and in original SSD
Vehicle attitude classification task is added on network model, wherein the vehicle attitude classification task and vehicle detection task are arranged side by side；
Multitask loss structure module, for default weight coefficient to be multiplied by the loss for the vehicle attitude classification task being added
Afterwards, it is added to form multitask loss with the loss of the vehicle detection task；
Loss function determining module, for the damage using focal loss losses as the vehicle attitude classification task being added
It loses, and focal loss is lost into the Classification Loss as the vehicle target in the loss of the vehicle detection task；
Joint training module, for building the vehicle detection network based on SSD that the vehicle attitude classification task is added, according to
The vehicle detection network generates the candidate target frame and each candidate target of fixed size on the characteristic pattern of different layers
The possibility for including vehicle in frame arranges all candidate target frames according to the confidence level that each candidate target frame includes vehicle
Sequence, and the low negative sample of confidence level is removed, so that the ratio of positive negative sample is maintained at preset ratio, according to the positive negative sample of reservation
Calculate the multitask loss, then backpropagation, repetitive exercise, until reaching maximum iteration obtains target detection mould
Type, to carry out vehicle detection to picture to be detected by the target detection model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810311110.8A CN108596053B (en) | 2018-04-09 | 2018-04-09 | Vehicle detection method and system based on SSD and vehicle posture classification |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810311110.8A CN108596053B (en) | 2018-04-09 | 2018-04-09 | Vehicle detection method and system based on SSD and vehicle posture classification |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108596053A true CN108596053A (en) | 2018-09-28 |
| CN108596053B CN108596053B (en) | 2020-06-02 |
Family
ID=63621165
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810311110.8A Expired - Fee Related CN108596053B (en) | 2018-04-09 | 2018-04-09 | Vehicle detection method and system based on SSD and vehicle posture classification |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108596053B (en) |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109326124A (en) * | 2018-10-17 | 2019-02-12 | 江西洪都航空工业集团有限责任公司 | A kind of urban environment based on machine vision parks cars Activity recognition system |
| CN109345435A (en) * | 2018-12-07 | 2019-02-15 | 山东晴天环保科技有限公司 | Occupy-street-exploit managing device and method |
| CN109409252A (en) * | 2018-10-09 | 2019-03-01 | 杭州电子科技大学 | A kind of traffic multi-target detection method based on modified SSD network |
| CN109447171A (en) * | 2018-11-05 | 2019-03-08 | 电子科技大学 | A kind of vehicle attitude classification method based on deep learning |
| CN109472214A (en) * | 2018-10-17 | 2019-03-15 | 福州大学 | One kind is taken photo by plane foreign matter image real-time detection method based on deep learning |
| CN109544522A (en) * | 2018-11-12 | 2019-03-29 | 北京科技大学 | A kind of Surface Defects in Steel Plate detection method and system |
| CN109766769A (en) * | 2018-12-18 | 2019-05-17 | 四川大学 | A kind of road target detection recognition method based on monocular vision and deep learning |
| CN109800778A (en) * | 2018-12-03 | 2019-05-24 | 浙江工业大学 | A kind of Faster RCNN object detection method for dividing sample to excavate based on hardly possible |
| CN109829469A (en) * | 2018-11-08 | 2019-05-31 | 电子科技大学 | A kind of vehicle checking method based on deep learning |
| CN109859190A (en) * | 2019-01-31 | 2019-06-07 | 北京工业大学 | A kind of target area detection method based on deep learning |
| CN110021164A (en) * | 2019-03-02 | 2019-07-16 | 合肥学院 | Net based on travel time data about bus or train route net occupation rate analysis method |
| CN110059544A (en) * | 2019-03-07 | 2019-07-26 | 华中科技大学 | A kind of pedestrian detection method and system based on road scene |
| CN110097055A (en) * | 2019-04-29 | 2019-08-06 | 上海工程技术大学 | A kind of vehicle attitude detection method and system based on grid convolutional neural networks |
| CN110163081A (en) * | 2019-04-02 | 2019-08-23 | 宜通世纪物联网研究院(广州)有限公司 | SSD-based real-time regional intrusion detection method, system and storage medium |
| CN110263679A (en) * | 2019-06-03 | 2019-09-20 | 西安交通大学 | A kind of fine granularity vehicle checking method based on deep neural network |
| CN110335319A (en) * | 2019-06-26 | 2019-10-15 | 华中科技大学 | Camera positioning and the map reconstruction method and system of a kind of semantics-driven |
| CN110379464A (en) * | 2019-07-29 | 2019-10-25 | 桂林电子科技大学 | The prediction technique of DNA transcription terminator in a kind of bacterium |
| CN110458225A (en) * | 2019-08-08 | 2019-11-15 | 北京深醒科技有限公司 | A kind of vehicle detection and posture are classified joint recognition methods |
| CN111461128A (en) * | 2020-03-31 | 2020-07-28 | 北京爱笔科技有限公司 | License plate recognition method and device |
| CN112016532A (en) * | 2020-10-22 | 2020-12-01 | 腾讯科技（深圳）有限公司 | Vehicle detection method and device |
| CN113128444A (en) * | 2021-04-28 | 2021-07-16 | 奇瑞汽车股份有限公司 | Loss function obtaining method, computer readable storage medium and electronic device |
| CN113792733A (en) * | 2021-09-17 | 2021-12-14 | 平安科技（深圳）有限公司 | Vehicle component detection method, system, electronic device and storage medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107578091A (en) * | 2017-08-30 | 2018-01-12 | 电子科技大学 | A kind of Pedestrians and vehicles real-time detection method based on lightweight depth network |
| CN107590438A (en) * | 2017-08-16 | 2018-01-16 | 中国地质大学（武汉） | A kind of intelligent auxiliary driving method and system |
-
2018
- 2018-04-09 CN CN201810311110.8A patent/CN108596053B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107590438A (en) * | 2017-08-16 | 2018-01-16 | 中国地质大学（武汉） | A kind of intelligent auxiliary driving method and system |
| CN107578091A (en) * | 2017-08-30 | 2018-01-12 | 电子科技大学 | A kind of Pedestrians and vehicles real-time detection method based on lightweight depth network |
Non-Patent Citations (3)
| Title |
|---|
| TSUNG-YI LIN 等: "Focal Loss for Dense Object Detection", 《THE IEEE ICCV 2017》 * |
| WEI LIU 等: "SSD: Single Shot MultiBox Detector", 《EUROPEAN CONFERENCE ON COMPUTER VISION》 * |
| 王林 等: "Faster R-CNN模型在车辆检测中的应用", 《计算机应用》 * |
Cited By (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109409252A (en) * | 2018-10-09 | 2019-03-01 | 杭州电子科技大学 | A kind of traffic multi-target detection method based on modified SSD network |
| CN109472214A (en) * | 2018-10-17 | 2019-03-15 | 福州大学 | One kind is taken photo by plane foreign matter image real-time detection method based on deep learning |
| CN109326124A (en) * | 2018-10-17 | 2019-02-12 | 江西洪都航空工业集团有限责任公司 | A kind of urban environment based on machine vision parks cars Activity recognition system |
| CN109447171A (en) * | 2018-11-05 | 2019-03-08 | 电子科技大学 | A kind of vehicle attitude classification method based on deep learning |
| CN109829469A (en) * | 2018-11-08 | 2019-05-31 | 电子科技大学 | A kind of vehicle checking method based on deep learning |
| CN109544522A (en) * | 2018-11-12 | 2019-03-29 | 北京科技大学 | A kind of Surface Defects in Steel Plate detection method and system |
| CN109800778A (en) * | 2018-12-03 | 2019-05-24 | 浙江工业大学 | A kind of Faster RCNN object detection method for dividing sample to excavate based on hardly possible |
| CN109345435A (en) * | 2018-12-07 | 2019-02-15 | 山东晴天环保科技有限公司 | Occupy-street-exploit managing device and method |
| CN109766769A (en) * | 2018-12-18 | 2019-05-17 | 四川大学 | A kind of road target detection recognition method based on monocular vision and deep learning |
| CN109859190A (en) * | 2019-01-31 | 2019-06-07 | 北京工业大学 | A kind of target area detection method based on deep learning |
| CN110021164A (en) * | 2019-03-02 | 2019-07-16 | 合肥学院 | Net based on travel time data about bus or train route net occupation rate analysis method |
| CN110059544A (en) * | 2019-03-07 | 2019-07-26 | 华中科技大学 | A kind of pedestrian detection method and system based on road scene |
| CN110163081A (en) * | 2019-04-02 | 2019-08-23 | 宜通世纪物联网研究院(广州)有限公司 | SSD-based real-time regional intrusion detection method, system and storage medium |
| CN110097055A (en) * | 2019-04-29 | 2019-08-06 | 上海工程技术大学 | A kind of vehicle attitude detection method and system based on grid convolutional neural networks |
| CN110263679A (en) * | 2019-06-03 | 2019-09-20 | 西安交通大学 | A kind of fine granularity vehicle checking method based on deep neural network |
| CN110335319A (en) * | 2019-06-26 | 2019-10-15 | 华中科技大学 | Camera positioning and the map reconstruction method and system of a kind of semantics-driven |
| CN110335319B (en) * | 2019-06-26 | 2022-03-18 | 华中科技大学 | Semantic-driven camera positioning and map reconstruction method and system |
| CN110379464A (en) * | 2019-07-29 | 2019-10-25 | 桂林电子科技大学 | The prediction technique of DNA transcription terminator in a kind of bacterium |
| CN110458225A (en) * | 2019-08-08 | 2019-11-15 | 北京深醒科技有限公司 | A kind of vehicle detection and posture are classified joint recognition methods |
| CN111461128A (en) * | 2020-03-31 | 2020-07-28 | 北京爱笔科技有限公司 | License plate recognition method and device |
| CN112016532A (en) * | 2020-10-22 | 2020-12-01 | 腾讯科技（深圳）有限公司 | Vehicle detection method and device |
| CN112016532B (en) * | 2020-10-22 | 2021-02-05 | 腾讯科技（深圳）有限公司 | Vehicle detection method and device |
| CN113128444A (en) * | 2021-04-28 | 2021-07-16 | 奇瑞汽车股份有限公司 | Loss function obtaining method, computer readable storage medium and electronic device |
| CN113128444B (en) * | 2021-04-28 | 2023-02-03 | 奇瑞汽车股份有限公司 | Loss function obtaining method, computer readable storage medium and electronic device |
| CN113792733A (en) * | 2021-09-17 | 2021-12-14 | 平安科技（深圳）有限公司 | Vehicle component detection method, system, electronic device and storage medium |
| CN113792733B (en) * | 2021-09-17 | 2023-07-21 | 平安科技（深圳）有限公司 | Vehicle part detection method, system, electronic device and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108596053B (en) | 2020-06-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108596053A (en) | A kind of vehicle checking method and system based on SSD and vehicle attitude classification | |
| CN109919108B (en) | Remote sensing image rapid target detection method based on deep hash auxiliary network | |
| Liu et al. | YOLO-extract: Improved YOLOv5 for aircraft object detection in remote sensing images | |
| CN110348437B (en) | Target detection method based on weak supervised learning and occlusion perception | |
| CN110348384B (en) | Small target vehicle attribute identification method based on feature fusion | |
| CN111160249A (en) | Multi-class target detection method of optical remote sensing image based on cross-scale feature fusion | |
| CN111091105A (en) | Remote sensing image target detection method based on new frame regression loss function | |
| CN110532859A (en) | Remote Sensing Target detection method based on depth evolution beta pruning convolution net | |
| CN104680542B (en) | Remote sensing image variation detection method based on on-line study | |
| CN110991311A (en) | Target detection method based on dense connection deep network | |
| CN109492596B (en) | Pedestrian detection method and system based on K-means clustering and regional recommendation network | |
| CN110399884A (en) | A kind of adaptive anchor frame model vehicle checking method of Fusion Features | |
| CN111489301A (en) | Image defogging method based on image depth information guide for migration learning | |
| CN110569782A (en) | Target detection method based on deep learning | |
| CN110826379A (en) | Target detection method based on feature multiplexing and YOLOv3 | |
| CN110716792B (en) | Target detector and construction method and application thereof | |
| CN113378763A (en) | SAR image-oriented target automatic detection method and system | |
| CN110263731B (en) | Single step human face detection system | |
| Wang et al. | Deep learning model for target detection in remote sensing images fusing multilevel features | |
| Wang et al. | Research on vehicle detection based on faster R-CNN for UAV images | |
| CN111368625B (en) | Pedestrian target detection method based on cascade optimization | |
| CN113408564A (en) | Graph processing method, network training method, device, equipment and storage medium | |
| CN114743151B (en) | Lightweight anti-interference flame detection method based on improved YOLOv-tiny | |
| CN116071704A (en) | Dual-mode network-based multi-data set X-ray security inspection image target detection method | |
| CN117173759A (en) | Face detection method based on multi-scale feature fusion |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20200602 |