JP7474689B2 - Object detection method and object detection device - Google Patents

Description

The present invention relates to an object detection method and an object detection device.

There are known inventions that use cameras and radar to recognize objects around a vehicle (Patent Document 1). The invention described in Patent Document 1 uses the distance to an object detected by the camera and radar to determine whether the object is the same object.

JP 2015-31607 A

However, the invention described in Patent Document 1 only evaluates the correspondence between points, so when a large object such as a bus is observed by multiple sensors, it is difficult to determine the correspondence between the detection results, including the size and shape, obtained from each sensor.

The present invention was made in consideration of the above problems, and its purpose is to provide an object detection method and an object detection device that can accurately determine the identity of an object using multiple sensors.

An object detection method according to one aspect of the present invention calculates a first polygonal shape that is a polygonal approximation of the object shape detected by a first sensor in a top view, calculates a second polygonal shape that is a polygonal approximation of the object shape detected by a second sensor in a top view, and determines whether the object detected by the first sensor and the object detected by the second sensor are the same object based on a first reference point that is a point at a predetermined position on a first side that has the largest intersection angle between the first sensor and an end point of the side among the sides that make up the first polygonal shape, and a second reference point that is a point at a predetermined position on a second side that has the largest intersection angle between the second sensor and an end point of the side among the sides that make up the second polygonal shape.

The present invention makes it possible to accurately determine the identity of objects using multiple sensors.

FIG. 1 is a configuration diagram of an object detection device 1 according to an embodiment of the present invention. FIG. 2 is a flowchart illustrating an example of the operation of the object detection device 1. FIG. 3 is a diagram for explaining an example of a method for determining a reference side. FIG. 4 is a diagram for explaining an example of a method for determining an observation reference point. FIG. 5 is a diagram illustrating an example of the observation error distribution. FIG. 6 is a diagram illustrating an example of a calculation delay error distribution. FIG. 7 is a diagram for explaining an example of a method for combining the observation error distribution and the calculation delay error distribution. FIG. 8 is a diagram illustrating an example of distance evaluation. FIG. 9 is a diagram illustrating an example of reference point allocation. FIG. 10 is a diagram illustrating an example of a representative observation vector.

Embodiments of the present invention will be described below with reference to the drawings. In the description of the drawings, identical parts are given the same reference numerals and the description will be omitted.

An example of the configuration of the object detection device 1 will be described with reference to FIG. 1. As shown in FIG. 1, the object detection device 1 includes a camera 10, a radar 11, a lidar 12, a controller 20, and a storage device 13.

The camera 10 has an imaging element such as a charge-coupled device (CCD) or a complementary metal oxide semiconductor (CMOS). The camera 10 detects objects around the vehicle (pedestrians, bicycles, motorcycles, other vehicles, etc.) and information about the vehicle's surroundings (division lines, traffic lights, signs, pedestrian crossings, intersections, etc.). The camera 10 outputs the captured images to the controller 20. The camera 10 may be a monocular camera or a stereo camera.

The radar 11 emits radio waves (emission waves) to objects around the vehicle and measures the reflected waves to measure the distance and direction to the objects. The radar 11 outputs the measured data to the controller 20. There are no particular limitations on the type of radar 11, but examples include millimeter wave radar.

LIDAR 12 (Laser Imaging Detection and Ranging) measures the distance and direction to objects around the vehicle, just like radar 11. LIDAR 12 differs from radar 11 in that radar 11 uses radio waves as its emitted waves, whereas LIDAR 12 uses light as its emitted waves. LIDAR 12 emits light to objects around the vehicle and measures the reflected light to measure the distance and direction to the objects and recognize the shape of the objects. LIDAR 12 outputs the measured data to controller 20.

The camera 10, radar 11, and lidar 12 each detect objects using a different method. The camera 10, radar 11, and lidar 12 are installed on the vehicle, but the installation locations are not particularly limited. There is also no limit to the number of installations. Only one of each may be installed, or multiple of each may be installed.

The storage device 13 is composed of a hard disk drive (HDD), a solid state drive (SSD), etc. The storage device 13 stores a database 14. The database 14 stores data on the observation errors and calculation delay errors of the camera 10, the radar 11, and the lidar 12.

The controller 20 is a general-purpose microcomputer equipped with a CPU (Central Processing Unit), memory, and input/output units. A computer program for functioning as the object detection device 1 is installed in the microcomputer. By executing the computer program, the microcomputer functions as the multiple information processing circuits equipped in the object detection device 1. Note that here, an example is shown in which the multiple information processing circuits equipped in the object detection device 1 are realized by software, but it is of course possible to configure the information processing circuits by preparing dedicated hardware for executing each information process shown below. The multiple information processing circuits may also be configured by individual hardware. The controller 20 includes a reference side determination unit 21, a reference point determination unit 22, an error distribution determination unit 23, a distance calculation unit 24, an allocation unit 25, and a state update unit 26 as an example of the multiple information processing circuits.

The reference edge determination unit 21 acquires data related to object detection from the camera 10, radar 11, and lidar 12. This data includes the size and shape of the observed object. The reference edge determination unit 21 determines the reference edge of the object using the data related to the size and shape of the object. The reference edge determination unit 21 outputs information indicating the determined reference edge to the reference point determination unit 22.

The reference point determination unit 22 uses the reference edge acquired from the reference edge determination unit 21 to set one observation reference point on this reference edge. The reference point determination unit 22 outputs information indicating the set reference point to the error distribution determination unit 23.

The error distribution determination unit 23 determines the error distribution of the observation reference point using the observation reference point acquired from the reference point determination unit 22 and data related to the observation error and calculation delay error stored in the database 14. The error distribution determination unit 23 also determines the error distribution of the target reference point using the average and covariance of the current state estimation result estimated one time ago acquired from the state update unit 26.

The distance calculation unit 24 evaluates the distance between the observation reference point including the error distribution obtained from the error distribution determination unit 23 and the target reference point including the error distribution.

The allocation unit 25 associates the observation results with the targets based on the distance evaluation results between the observation reference points and the target reference points obtained from the distance calculation unit 24. If there is no target to which an assignment can be made, the allocation unit 25 adds a new target, and if no observation results have been assigned to the target, it deletes the target. Furthermore, the allocation unit 25 integrates the observation information assigned to the target according to the reliability (priority) of each sensor, and determines the representative observation vector of the target.

The state update unit 26 updates the state vector of the target based on the representative observation vector of the target obtained from the allocation unit 25.

Next, an example of the operation of the object detection device 1 will be described with reference to Figures 2 to 10.

In step S101 shown in FIG. 2, objects around the vehicle are detected by multiple different sensors, namely, the camera 10, the radar 11, and the lidar 12. The object detection results obtained from each sensor are expressed as an observation vector consisting of the object's position, attitude, first derivative of the position (velocity), first derivative of the attitude (angular velocity), and size (length and width of a rectangle). However, not all sensors can necessarily measure all information, and even for information that can be observed, there are differences in reliability (measurement error distribution) due to differences in measurement principles.

For example, when the camera 10 is a monocular camera, the detection results obtained from the monocular camera accurately measure the width of the object, but the accuracy of the depth of the object is low. Also, the positional accuracy in the line of sight is low, but the positional accuracy in the horizontal direction (azimuth angle) is high.

The detection results obtained by the radar 11 using a millimeter wave radar are characterized by high positional accuracy in the line of sight (the direction in which the emitted wave is emitted) and low positional accuracy in the horizontal direction. Furthermore, the detection results obtained by the lidar 12 are characterized by accurate measurement of the width and depth of an object, while also having high positional accuracy in the line of sight and horizontal directions. However, because it is not always possible to observe the width and depth of an object depending on the relative positions of the sensor and the object being observed, the observation data for the object changes successively.

Even if the observation data about the object changes sequentially depending on the relative positions of the sensor and the object being observed, in order to uniformly evaluate the correspondence between the observation results and the target while taking into account the difference in measurement accuracy between sensors with different measurement principles, the observation reference point for the observation results is determined and an observation error distribution is assigned in steps S103, 105, and 107.

In step S103, the reference side determination unit 21 determines the reference side based on the detection results including the size and shape (rectangle) of the object obtained from the camera 10, the radar 11, and the lidar 12. The reference side will be described with reference to FIG. 3.

In the example shown in FIG. 3, the two sides on the sensor side of the rectangular object area 30 seen from the camera 10, the radar 11, and the lidar 12 are targeted in determining the reference side. The side with the larger crossing angle with respect to the line (line of sight direction) connecting the sensor point and the end points of the rectangular object area 30 is set as the first reference side L1, and the side with the smaller crossing angle is set as the second reference side L2. In this case, the reference side determination unit 21 determines the first reference side L1 as the reference side. θ1 is the angle between the line of sight direction and the extension line 40 of the first reference side L1 at the end point (point on the lower right) of the rectangular object area 30. Also, θ2 is the angle between the line of sight direction and the extension line 41 of the second reference side L2 at the end point (point on the upper left) of the rectangular object area 30. θ1>θ2. The reason why the two sides closer to the sensor side are targeted is because the sensor accuracy is higher for the two sides closer to the sensor side than for the two sides farther from the sensor side. The reference side may be expressed as a side related to a part having a larger area as seen from the sensor.

In step S105 shown in FIG. 2, the reference point determination unit 22 sets an arbitrary point on the first reference side L1 determined by the reference side determination unit 21 as the observation reference point. In this embodiment, as shown in FIG. 4, the midpoint of the first reference side L1 is determined as the observation reference point 50. However, as described above, the observation reference point is not limited to the midpoint, and may be any point on the first reference side L1. The observation reference point is determined using the average of the observation vectors acquired from each sensor or the average and covariance of the observation vectors.

In step S107 shown in FIG. 2, the error distribution determination unit 23 determines the error distribution of the observation reference point using the observation reference point 50 acquired from the reference point determination unit 22 and the observation error distribution and calculation delay error distribution acquired from the database 14.

Here, the observation error distribution is a representation of the observation error caused by the different observation principles for each sensor, expressed by a covariance matrix. A covariance matrix for the observation error for each sensor is set at the observation reference point, and is set as the allowable error for each observation reference point. For example, as shown in FIG. 5, an error distribution is set in the direction of a straight line (line of sight) connecting the sensor point and the observation reference point, and in the direction perpendicular to this line of sight (horizontal direction). This determines the observation error distribution. In FIG. 5, reference numeral 50 indicates the observation reference point of the lidar 12, and reference numeral 51 indicates the observation reference point of the camera 10. Reference numeral 60 indicates a polygonal shape obtained by approximating the object shape detected by the lidar 12 from a top view, and reference numeral 61 indicates a polygonal shape obtained by approximating the object shape detected by the camera 10 from a top view. Such polygonal shapes are calculated by the controller 20.

In Figure 5, symbol L3 indicates the line of sight vector as seen by the LIDAR 12, and symbol L4 indicates the line of sight vector as seen by the camera 10. Symbol 50a indicates the observation error distribution of the LIDAR 12, and symbol 51a indicates the observation error distribution of the camera 10. In the example shown in Figure 5, the observation error distribution indicates that the positional accuracy of the line of sight by the camera 10 is low and the positional accuracy in the horizontal direction is high, while the positional accuracy of the LIDAR 12 is high in both the line of sight and horizontal directions.

On the other hand, the calculation delay error distribution is a representation of the observation error caused by the different calculation processing cycles for each sensor, expressed by a covariance matrix. A covariance matrix related to the calculation delay error for each sensor is set at the observation reference point, and this is set as the allowable error for each observation reference point. For example, as shown in Figure 6, an error distribution is set in the velocity vector direction and the orthogonal direction for each observation. This determines the calculation delay error distribution. Reference L5 in Figure 6 indicates the velocity vector for the LIDAR 12, and reference L6 indicates the velocity vector for the camera 10. Reference 50b indicates the calculation delay error distribution for the LIDAR 12, and reference 51b indicates the calculation delay error distribution for the camera 10.

The final error distribution of the observation reference point can be determined by combining the observation error distribution and the calculation delay error distribution according to the additivity of the normal distribution as shown in Figure 7. In Figure 7, the error distribution 51c of the observation reference point 51 is determined by combining the observation error distribution 51a and the calculation delay error distribution 51b. Note that the error distribution 51c indicates the error distribution of the camera 10, and although not shown, the error distributions of the radar 11 and lidar 12 are also determined by a similar combination.

By expressing the existence probability distribution of each observation point as a normal distribution in this way, it is possible to probabilistically express the existence of the observation reference point, reflecting the observation error due to the measurement principle of each sensor and the calculation delay due to the magnitude of the calculation load. For example, suppose that the normal probability distribution reflecting the observation error of an arbitrary observation point 1 is expressed by Equation 1.

where μ ₁ is the mean of the observation points and Σ ₁₁ is the covariance matrix of the observation points.

Let us assume that the normal probability distribution reflecting the calculation delay error at an arbitrary observation point 1 is expressed by Equation 2.

Here, Σ ₁₂ is the covariance matrix of the observation points.

The final normal probability distribution for observation point 1 is expressed by equation 3.

In addition, if state estimation of the observation results is performed within each sensor and a posterior probability distribution of the state estimation results is obtained, this value may be used as the existence probability distribution of each observation point. State estimation here means tracking.

The observation error distribution stored in database 14 is obtained by acquiring the measurement values of each sensor in an environment where the correct value of the observation target is known, and calculating the error between the correct value and the measurement value. Similarly, the calculation delay error distribution is obtained by measuring the time required for measurement and calculation of each sensor in an environment where the correct value of the observation target is known.

Furthermore, in step S107, the error distribution determination unit 23 determines the target reference point and the error distribution using the average and covariance of the state estimation result for the current time calculated one time ago, obtained from the state update unit 26. The target reference point will be explained. One time ago from the present is defined as "time t-1", and the present is defined as "time t". In this case, in this embodiment, the controller 20 predicts the state quantity of the future time t at the time t-1. Such prediction is made based on the viewpoint of how the position, attitude, speed, etc. at the time t-1 will change one time later. The predicted observation reference point at this time is the "target reference point". In other words, the target reference point is different from the observation reference point based on the current observation results of the camera 10, the radar 11, and the lidar 12.

In step S109, the distance calculation unit 24 performs a distance evaluation between the observation reference point including the error distribution acquired from the error distribution determination unit 23 and the target reference point including the error distribution. This distance evaluation will be described with reference to Figs. 8 to 9. Reference d1 in Fig. ₈ indicates the distance between the target reference point 71 and the observation reference point 51. Reference _d2 in Fig. 8 indicates the distance between the target reference point 71 and the observation reference point 55. When the observation reference point and the target reference point both include error distributions, the evaluation is of the distance between normal distributions. Therefore, if the probability distribution of the observation reference point is p(x) and the probability distribution of the target reference point is q(x), then, for example, the KL divergence can be used as the distance measure (Equation 4).

Note that the distance between reference points does not necessarily have to be the distance between normal distributions. For example, if either the observation reference point or the target reference point is represented by the average of a normal distribution, the distance will be the distance between the point and the normal distribution. Therefore, if the observation reference point represented by a point is x _j and the normal distribution to which the distance is to be calculated is N(μ _i , Σ _i ), then, for example, the Mahalanobis distance can be used as the distance measure (Equation 5).

In addition, when both the observation reference point and the target reference point are expressed by the mean of a normal distribution, the distance is between points. Therefore, if the observation reference points expressed as points are x _i and x _j , respectively, the squared error, for example, can be used as the distance measure (Equation 6).

In step S111, the allocation unit 25 associates the observation results with the target based on the distance evaluation result between the observation reference point and the target reference point obtained from the distance calculation unit 24. For the observation reference points obtained from each sensor, a process of allocating an observation reference point to the target if the distance from the target reference point to the observation reference point that is the shortest is equal to or less than a threshold. For example, when considering allocating observation reference point i to target reference point j, the allocation unit 25 selects observation reference point i so as to satisfy the condition expressed in the following formula 7.

Specifically, referring to FIG. 9, the observation reference point with the smallest distance from the target reference point 71 is the observation reference point 51. Since the distance between the target reference point 71 and the observation reference point 51 is equal to or less than the threshold, the observation reference point 51 is assigned to the target 70. In the example shown in FIG. 9, the case where the observation reference point 51 of the camera 10 is assigned has been described, but the observation reference points of the radar 11 and the lidar 12 are assigned in a similar manner. When the observation reference points of each sensor are assigned to the target reference point in this manner, the controller 20 determines that the objects detected by each sensor are the same object. Note that FIG. 9 shows that another observation reference point 55 of the camera 10 is assigned to the target reference point 81.

In this way, two or more observation reference points obtained from the same sensor (camera 10 in Figure 9) are not assigned to one target reference point, and only one observation result is assigned from the same sensor at most. Also, if there is no target to assign, a new target is added with the value of the observation reference point as the initial value, and if a certain amount of time has passed without any observation result being assigned to the target, the target is deleted.

Furthermore, in step S111, for the multiple observation results assigned to the target, the observation information is integrated according to the reliability (priority) of each sensor to determine a representative observation vector of the target. The reliability of each sensor can be determined by evaluating the error distribution of the observation reference point assigned to the target and increasing the reliability of the sensor as the variance decreases. The representative observation vector will be described with reference to FIG. 10. Reference symbol L7 in FIG. 10 is the representative observation vector. Reference symbol 52 indicates the observation reference point of the radar 11, and reference symbol 62 indicates a polygonal shape that approximates the object shape detected by the radar 11 in a top view. As shown in FIG. 10, if the error distribution of the position in the line of sight of the camera 10 is larger than the error distribution of the position in the line of sight of the lidar 12 and the error distribution of the horizontal position of the camera 10 is smaller than the error distribution of the horizontal position of the lidar 12, the measurement result of the lidar 12 is used for the position in the line of sight, and the measurement result of the camera 10 is used for the horizontal position. In this way, by evaluating the error distribution of the observation reference points assigned to the target 70 and determining the priority, the observation information can be integrated and a representative observation vector L7 can be determined.

In step S113, the state update unit 26 updates the state vector of the target 70 based on the representative observation vector L7 of the target 70 acquired from the allocation unit 25. A known state space model such as an extended Kalman filter (EKF) or a particle filter is used for the state update.

(Action and Effect)
As described above, the object detection device 1 according to this embodiment provides the following advantageous effects.

The controller 20 calculates a polygonal shape that approximates the object shape detected by each sensor in a top view. If the sensor is a camera 10, the polygonal shape (first polygonal shape) is indicated by reference numeral 61 in FIG. 5. Alternatively, if the sensor is a lidar 12, the polygonal shape (second polygonal shape) is indicated by reference numeral 60 in FIG. 5. The controller 20 determines whether the object detected by the first sensor and the object detected by the second sensor are the same object based on a first reference point (e.g., observation reference point 51) that is a point at a predetermined position on a first side that has the largest intersection angle between the first sensor and the end point of the side among the sides that constitute the first polygonal shape, and a second reference point (e.g., observation reference point 50) that is a point at a predetermined position on a second side that has the largest intersection angle between the second sensor and the end point of the side among the sides that constitute the second polygonal shape. In this way, for detection results including the size and shape of the observed object, the surface that is most likely to be observed from each sensor is set as the reference edge, and a point on the reference edge is set as the reference point, so it is possible to accurately evaluate the correspondence between the observation information obtained from each sensor using information that can be observed even if the relative position of the observed object to each sensor changes. This makes it possible to accurately determine the identity of objects using multiple sensors.

The first sensor is, for example, a camera 10, which captures an image of a first predetermined range and detects an object present in the first predetermined range from the captured image. The second sensor is, for example, a radar 11 or a lidar 12, which emits an outgoing wave into a second predetermined range that overlaps with the first predetermined range and detects an object present in the second predetermined range based on the reflected wave of the outgoing wave. Note that the identity of an object may be determined using all of the sensors, the camera 10, the radar 11, and the lidar 12.

The controller 20 determines the observation reference point using the average of the observation vectors obtained from each sensor or the average and covariance of the observation vectors. The controller 20 determines the error distribution of the observation reference point by combining the observation error distribution and the calculation delay error distribution. This makes it possible to set a covariance matrix that reflects the observation error characteristics and calculation delay characteristics of each sensor at each observation point, making it possible to accurately evaluate the correspondence between multiple observation results even between observation points obtained from sensors with different characteristics.

The controller 20 sets a covariance matrix relating to the observation error for each sensor at the observation reference point.
The controller 20 also sets an error distribution in the direction of a line connecting the sensor point and the observation reference point, and in a direction perpendicular to the line direction. Since a different error distribution can be set for each sensor with a different measurement principle, such as a large error in the line of sight (line direction) and a small error in the azimuth angle direction for the camera 10, a small error in the line of sight direction and a large error in the azimuth angle direction for the radar 11, and a small error in the line of sight direction and an large error in the azimuth angle direction for the lidar 12, it is possible to evaluate the correspondence between multiple observation results within a common framework even if the number of sensor types increases, without relying on a single evaluation method.

The controller 20 sets a covariance matrix for the calculation delay error for each sensor at the observation reference point. The controller 20 also sets an error distribution in the velocity vector direction of the sensor point and in a direction perpendicular to the velocity vector direction. Since it is possible to individually set the error distribution caused by the calculation delay of sensors that perform object detection with different calculation periods, such as the camera 10, radar 11, and lidar 12, it is possible to evaluate the correspondence between multiple observation results within a common framework even if the number of sensor types increases, without relying on a single evaluation method.

The controller 20 uses a target reference point that indicates the current observation reference point estimated one time ago and the observation reference point to determine whether the object detected by the first sensor and the object detected by the second sensor are the same object. This makes it possible to accurately determine the identity of objects using multiple sensors.

The controller estimates the target reference point using the average of the state estimation results or the average and covariance of the state estimation results. The controller 20 determines the average and covariance of the target reference point using an extended Kalman filter, a particle filter, or the like. For targets that combine observation results obtained from multiple sensors, a covariance matrix can be set for the target reference point using a posterior distribution that reflects the reliability of the sequential state estimation results, so that the distance of the observation reference point to the target can be evaluated with high accuracy. This allows for accurate grouping and tracking of observation information while adding or deleting targets at appropriate times.

The controller 20 calculates the distance between the observation reference point and the target reference point using their respective averages or both the average and covariance. The controller 20 may also calculate the distance between the observation reference point and the target reference point using a normal distribution and the KL divergence of the normal distribution. The controller 20 may also calculate the distance between the observation reference point and the target reference point using the Mahalanobis distance of the observation reference point and the normal distribution. The controller 20 may calculate the distance between the observation reference point and the target reference point using the squared error between the observation reference point and the observation reference point. In addition to evaluating the distance between the observation reference point and the observation reference point, the controller 20 can also perform a distance evaluation that reflects the allowable error of the observation reference point and the allowable error of the target reference point, so that the controller 20 can evaluate the distance between the observation reference point and the target reference point while reflecting the observation error of the sensor and the reliability of tracking.

The controller 20 integrates the observation information according to a priority determined by evaluating the error distribution of the observation reference point for multiple observation results assigned to the target, and determines a representative observation vector for the target. Since it becomes possible to selectively integrate observation information with small errors from multiple types of sensors in the error distribution of the observation reference point, a highly reliable observation vector can be used as the representative observation vector for the target.

Each of the functions described in the above embodiments may be implemented by one or more processing circuits. Processing circuits include programmed processing devices, such as processors that include electrical circuitry. Processing circuits also include devices, such as application specific integrated circuits (ASICs) or circuit components, arranged to perform the described functions.

As described above, an embodiment of the present invention has been described, but the descriptions and drawings that form part of this disclosure should not be understood as limiting this invention. Various alternative embodiments, examples, and operating techniques will become apparent to those skilled in the art from this disclosure.

Reference Signs List 1: Object detection device 10: Camera 11: Radar 12: Lidar 13: Storage device 14: Database 20: Controller 21: Reference side determination unit 22: Reference point determination unit 23: Error distribution determination unit 24: Distance calculation unit 25: Assignment unit 26: State update unit

Claims (17)

1. An object detection method comprising: a first sensor that detects an object around a vehicle; a second sensor that detects the object in a manner different from that of the first sensor; and a controller that processes data detected by the first sensor and the second sensor,
The controller:
calculating a first polygonal shape by approximating the object shape detected by the first sensor as a polygon in a top view;
calculating a second polygonal shape by approximating the object shape detected by the second sensor as a polygon in a top view;
An object detection method characterized by determining whether the object detected by the first sensor and the object detected by the second sensor are the same object based on a first reference point which is a point at a predetermined position on a first side among the sides that constitute the first polygonal shape, which has the largest intersection angle with a line segment connecting the first sensor and an end point of the side, and a second reference point which is a point at a predetermined position on a second side among the sides that constitute the second polygonal shape, which has the largest intersection angle with a line segment connecting the second sensor and an end point of the side. The first sensor captures an image of a first predetermined range and detects an object present within the first predetermined range from the captured image;
The object detection method according to claim 1 , wherein the second sensor emits an outgoing wave into a second predetermined range that overlaps with the first predetermined range, and detects an object present in the second predetermined range based on a reflected wave of the outgoing wave. The object detection method according to claim 1 or 2, characterized in that the controller determines the first reference point and the second reference point using an average of observation vectors acquired from each sensor or the average and covariance of the observation vectors. The object detection method according to any one of claims 1 to 3, characterized in that the controller determines the error distribution of the first reference point and the second reference point by combining an observation error distribution and a calculation delay error distribution. The object detection method according to claim 4 , wherein the controller sets a covariance matrix relating to an observation error for each sensor to the first reference point and the second reference point. The object detection method described in claim 5, characterized in that the controller sets an error distribution in a straight line direction connecting the sensor point and the first reference point and in a direction perpendicular to the straight line direction, and sets an error distribution in a straight line direction connecting the sensor point and the second reference point and in a direction perpendicular to the straight line direction. The object detection method according to any one of claims 4 to 6, characterized in that the controller sets a covariance matrix relating to a calculation delay error for each sensor to the first reference point and the second reference point. 8. The object detection method according to claim 7, wherein the controller sets an error distribution in a velocity vector direction of the sensor point and in a direction perpendicular to the velocity vector direction. The object detection method described in any one of claims 1 to 8, characterized in that the controller determines whether the object detected by the first sensor and the object detected by the second sensor are the same object using a target reference point indicating a current observation reference point estimated one time ago, the first reference point, and the second reference point. The method of claim 9 , wherein the controller estimates the target reference point using an average of state estimation results or an average and covariance of the state estimation results. The method of claim 9 , wherein the controller determines the mean and covariance of the target reference points using an extended Kalman filter or a particle filter. The object detection method according to any one of claims 9 to 11, characterized in that the controller calculates the distance between the first reference point and the target reference point, and the distance between the second reference point and the target reference point using their respective averages or averages and covariances. The object detection method according to any one of claims 9 to 11, characterized in that the controller calculates the distance between the first reference point and the target reference point, and the distance between the second reference point and the target reference point, using a normal distribution and a KL divergence of a normal distribution. The object detection method according to any one of claims 9 to 11, characterized in that the controller calculates the distance between the first reference point and the target reference point, and the distance between the second reference point and the target reference point, using a Mahalanobis distance between a reference point and a normal distribution. The object detection method according to any one of claims 9 to 11, characterized in that the controller calculates the distance between the first reference point and the target reference point, and the distance between the second reference point and the target reference point, using the squared error between the reference points. The controller:
Integrating the observation information according to a priority determined by evaluating the error distribution of the first reference point and the second reference point for a plurality of observation results assigned to a target;
The method according to any one of claims 1 to 15, further comprising determining a representative observation vector of the target. A first sensor for detecting objects around the vehicle;
a second sensor that detects the object in a manner different from that of the first sensor;
a controller for processing data detected by the first sensor and the second sensor;
The controller:
calculating a first polygonal shape by approximating the object shape detected by the first sensor as a polygon in a top view;
calculating a second polygonal shape by approximating the object shape detected by the second sensor as a polygon in a top view;
an object detection device which determines whether the object detected by the first sensor and the object detected by the second sensor are the same object based on a first reference point which is a point at a predetermined position on a first side which, among the sides constituting the first polygonal shape, forms the largest intersection angle with a line segment connecting the first sensor and an end point of the side, and a second reference point which is a point at a predetermined position on a second side which, among the sides constituting the second polygonal shape, forms the largest intersection angle with a line segment connecting the second sensor and an end point of the side.

Images