Fusion target recognition method based on UV disparity detection and YOLOv5

Abstract

The present invention belongs to the field of image processing and computer vision, and relates to a fusion target recognition method based on UV disparity detection and YOLOv5. This method fully integrates the improved UV disparity detection and YOLOv5 model, uses the improved UV disparity detection to roughly identify pedestrians and vehicles and non-standard obstacles on the road, inputs the UV detection results into the YOLOv5 deep learning model, and fuses the two results for target recognition to obtain a highly stable and robust detection effect, and the recognition range is not limited to the type of target, thereby realizing the fusion target recognition function of binocular stereo vision. The present invention can efficiently and quickly realize the detection of drivable areas on the road, the detection of non-standard obstacles, and the classification and recognition of targets. It is a fusion target recognition algorithm with high stability and high robustness that combines the advantages of traditional binocular detection and deep learning target detection.

Description

Technical Field

[0001] The present invention belongs to the fields of image processing and computer vision, and relates to a binocular target recognition and detection method based on stereo vision technology, and specifically to a fusion target recognition method based on UV disparity detection and YOLOv5. Background Art

[0002] Previous detection algorithms mostly relied on deep learning, utilizing deep learning techniques and neural network models to perform extensive learning and training of the target under specific circumstances to achieve corresponding target detection and recognition. However, target detection and recognition algorithms that rely on neural networks and data annotation also have many problems. For example, when recognizing targets with a small number of training samples, recognition accuracy and recognition ability can be significantly reduced. When recognizing unknown targets, recognition timeliness can lead to problems with the entire recognition result. These shortcomings have resulted in previous target recognition algorithms lacking the ability to adapt to multiple scenarios, as well as the robustness and stability of the algorithms, making them difficult to implement in practice.

[0003] However, binocular stereo vision is an important research area in computer vision. Binocular stereo vision is based on high-quality imaging of the three-dimensional world from different positions and angles using two or more image sensors. By simulating the stereoscopic imaging principle of the human eye, stereo disparity is calculated on the two-dimensional images received by the image sensors, thereby outputting disparity images. By determining the three-dimensional spatial position of the camera and combining it with the principle of triangulation, the disparity images can be used to calculate the three-dimensional coordinates of each object in three-dimensional space. Due to its advantages such as low cost, strong stability, and simple structure, binocular stereo vision is increasingly being used in fields such as target recognition.

[0004] Because binocular stereo vision calculates images solely from the two-dimensional images recognized by the camera in three-dimensional space, its recognition stability is easily affected by ambient lighting, object color, and other factors, leading to deviations in the calculated disparity image. Current object recognition algorithms, such as UV disparity calculation and Stixels, often suffer from incomplete object recognition and inaccurate ground detection. Therefore, a highly stable and robust fusion object recognition algorithm is needed that combines the advantages of traditional binocular detection and deep learning object detection. Summary of the Invention

[0005] This paper proposes a fused target recognition and detection method based on binocular stereo vision. This method fully integrates improved UV disparity detection and the YOLOv5 model. It uses the improved UV disparity detection to identify non-standard obstacles on the road, and uses the YOLOv5 deep learning model to detect people and vehicles on the road. The results of these two methods are then fused for target recognition, resulting in highly stable and robust detection results. Furthermore, the recognition range is not limited to the target type, thereby achieving the fused target recognition function of binocular stereo vision.

[0006] The technical solutions of the present invention are as follows:

[0007] The target recognition method based on UV disparity detection and YOLOv5 is as follows:

[0008] 1) Collect the dedistorted left and right images of the road scene using a binocular camera;

[0009] 2) Calculate the disparity based on the collected left and right images of the road scene to obtain the original disparity map;

[0010] 3) constructing a V-disparity map based on the original disparity map obtained in step 2);

[0011] 4) Divide the road surface according to the V-disparity map and determine the target detection area;

[0012] 5) Pixel association is performed between the target detection area determined in step 4) and the V disparity map obtained in step 3), disparity information of pixels below the road surface in the V disparity map is removed, the vertical line where the obstacle intersects the road surface is detected, and a new disparity map containing only disparity information above the road surface is constructed;

[0013] 6) Based on the new disparity map obtained in step 5), a U-disparity map is constructed and obstacle lines are detected;

[0014] 7) The straight lines detected in steps 5) and 6) are comprehensively verified on the left image of the road scene. Contour recognition is performed on pedestrians and vehicles among obstacles to distinguish between pedestrians, vehicles and non-standard obstacles, and the UV parallax target detection results are combined;

[0015] 8) Input the UV disparity target detection result obtained in step 7) into the YOLOv5 model for joint training detection, output the detection result and combine it with the corresponding original disparity map to determine the target's two-dimensional image coordinates and three-dimensional space coordinates.

[0016] Beneficial effects of the present invention:

[0017] This invention improves on traditional UV parallax detection and integrates the detection results of the YOLOv5 model, efficiently and quickly realizing the functions of drivable road area detection, non-standard obstacle detection, and target classification and recognition. It further solves the errors caused by external conditions such as temperature, humidity, and vibration in binocular camera recognition. Compared with traditional UV parallax detection, the fusion detection of the two methods not only has the characteristics of more recognition types and higher recognition accuracy, but also further optimizes the target recognition speed and greatly reduces the amount of data required for image calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of the pinhole camera model.

[0019] Figure 2 This is the correspondence diagram between the pixel coordinate system of the image and the physical coordinate system.

[0020] Figure 3 This is a diagram showing the relationship between the camera coordinate system in the X-axis direction and the image physical coordinate system.

[0021] Figure 4 This is the relationship diagram between the camera coordinate system in the Y-axis direction and the image physical coordinate system.

[0022] Figure 5 Schematic diagram of parallel binocular stereo vision.

[0023] Figure 6 It is the conversion diagram between image pixel coordinates and image physical coordinate system.

[0024] Figure 7 This is the principle diagram of stereo matching.

[0025] Figure 8 Schematic diagram of polar constraint.

[0026] Figure 9 Construct a schematic diagram for the V-disparity map.

[0027] Figure 10 This is the flow chart of the road surface detection algorithm.

[0028] Figure 11 Construct a schematic diagram for the U-disparity map.

[0029] Figure 12 Flowchart of the vehicle and pedestrian contour recognition algorithm.

[0030] Figure 13 This is a YOLOv5 combined with UV disparity detection model diagram.

[0031] Figure 14 Schematic diagram of YOLOv5 combined with UV disparity detection.

[0032] Figure 15The basic flow chart of the method of the present invention is shown in FIG. DETAILED DESCRIPTION

[0033] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0034] The basic process of the target recognition method based on UV parallax detection and YOLOv5 of the present invention is as follows: Figure 15 As shown, the details are as follows:

[0035] 1) The binocular camera collects the left and right images of the road scene after dedistortion. The imaging model involved is: a barrier with a small hole is added between the object and the film, which constitutes pinhole imaging. The mathematical model is called the pinhole camera model. The small hole in the barrier is called the aperture, and an inverted image is obtained on the film. The distance from the object to the film is Z, which is the depth, and the distance from the barrier to the film is F, which is the focal length. The schematic diagram of the pinhole camera model is shown below. Figure 1 shown.

[0036] The camera optical imaging process mainly involves the following coordinates and transformations:

[0037] a. Image pixel coordinate system: This represents the projection of objects in a real-world scene onto a two-dimensional image. The origin of the pixel coordinate system is the upper-left corner of the two-dimensional image. The horizontal axis (U axis) runs horizontally to the right along the image, and the vertical axis (V axis) runs vertically downward along the image. Coordinates are often expressed as (U, V), which represent the column and row numbers of the pixel in the image pixel coordinate system, respectively, and are measured in pixels.

[0038] b. Image physical coordinate system: Since the image pixel coordinate system only represents the number of columns and rows of image pixels in the pixel matrix of a two-dimensional image, and does not use physical units to represent pixel points, it is necessary to establish an image physical coordinate system represented by physical units. The image physical coordinate system is actually the surface of the camera sensor, that is, the image plane. The coordinate origin is defined as the intersection of the camera optical axis and the image plane. This point is generally located at the center of the image (U_0, V_0), but due to process manufacturing and other reasons, there will be an offset in actual situations. The horizontal axis (X-axis) is parallel to the U-axis, and the vertical axis (Y-axis) is parallel to the V-axis. The coordinates are usually expressed as (x, y) in millimeters (mm), such as Figure 2 shown.

[0039] c. Camera coordinate system: The camera coordinate system takes the optical center of the camera as the origin of the coordinate system. Its X-axis and Y-axis are parallel to the X-axis and Y-axis of the image physical coordinate system respectively. The optical axis of the camera is the Z-axis of the camera coordinate system. The unit is meter (m). The relationship between the camera coordinate system and the image physical coordinate system is as follows: Figure 3 and Figure 4 shown.

[0040] d. World coordinate system: The world coordinate system is an absolute coordinate system that describes the objective three-dimensional world. Since the camera is installed in the real three-dimensional space, the world coordinate system is needed to describe the camera's position in the objective three-dimensional world, and this coordinate system is used to describe the specific positions of other objects in the objective three-dimensional world.

[0041] 2) Calculate the disparity of the collected left and right images of the road scene to obtain a disparity map. The steps involved in the disparity calculation are as follows:

[0042] Schematic diagram of parallel binocular stereo vision Figure 5 shown.

[0043] The conversion relationship between the image pixel coordinate system and the camera coordinate system is:

[0044] The conversion between image pixel coordinates and image physical coordinate system is as follows: Figure 6 As shown, the origin of the image coordinate system is at the center of the imaging plane, and the origin of the pixel coordinate is at the upper left corner of the image, that is, the horizontal coordinate of the pixel coordinate system is the sum of half the image height and the horizontal coordinate of the image coordinate system, and the vertical coordinate of the pixel coordinate system is the sum of half the image width and the vertical coordinate of the image coordinate system.

[0045] After obtaining the disparity and other parameters of the pixel point, the coordinates of the point in the camera coordinate system can be further calculated. l -U r The disparity value at that point is obtained by simply calculating the internal parameters of the binocular camera, the baseline length, the image coordinates, and the disparity to calculate the coordinates of that spatial point in the camera coordinate system. The internal parameters of the binocular camera include a. the disparity value at infinity, b. the baseline length, c. the focal length, d. the pixel size (e) and the field of view (y) angle, and f. the camera's height above the ground.

[0046] The above data can be used to calculate stereo disparity. Disparity calculation requires knowing two projection points. However, in the actual calculation process, since we do not know which two points in the binocular camera are the corresponding points, we need to use stereo matching to find the corresponding points. Figure 7 shown.

[0047] Figure 7 In the figure, points P and Q are mapped to the same point on the image plane of the left camera OR. As long as the corresponding points of points P and Q on the image plane of the right camera OT are found, the depth can be estimated by triangulation. This process is called stereo matching. Finding corresponding points through stereo matching requires adding constraints, and the most commonly used constraint is the epipolar constraint, such as Figure 8 shown.

[0048] Figure 8In the figure, points PQ are mapped to the same point on the left camera image plane. The corresponding point on line PQ lies on line P'Q' on the right image plane. This line is the epipolar line of line PQ. This constraint is called an epipolar constraint. Methods for finding corresponding points can be roughly divided into three categories: regional stereo matching algorithms, feature-based cubic matching methods, and correlation methods. This invention uses the semi-global matching algorithm (SGM) within the regional stereo matching method.

[0049] 3) The disparity map obtained in step 2) is subjected to a V-disparity map construction, wherein the V-disparity map construction process of the disparity images involved is as follows.

[0050] Since the V disparity map is obtained by accumulating the number of pixels with the same disparity in each row in the V direction, the disparity map is scanned row by row from left to right along the V direction, and the vertical length of the disparity map is used as the vertical coordinate of the V disparity map, and the accumulated number of pixels with the same disparity value in each row is used as the intensity of each coordinate point of the V disparity map, thereby forming a "vertical coordinate-disparity value-number" map (i.e., V disparity map), as shown in the figure. Figure 9 shown.

[0051] 4) Divide the road surface according to the V disparity map and determine the target detection area. The process of determining the target detection area is as follows: Figure 10 As shown, the details are as follows:

[0052] 4.1) Determining the Horizon: The road surface area is mapped in the V disparity map, with the lower left corner as the origin. Disparity in the V direction decreases as the coordinate increases. Ideally, the disparity statistical point corresponding to the road surface area in the V disparity map is a continuous diagonal line segment, representing the road surface area. This diagonal line segment encompasses the vertical extent of the road surface area in the original image. Because the road surface area extends from the bottom of the image to the horizon, the horizon position is determined first. First, the number of rows where the road surface contour stops extending to the left is set as the row containing the horizon. The height Hn of the horizon row is determined from the input V disparity map. To reduce the algorithm's computational complexity, locations above the horizon are set as non-road areas. The disparity of rows above Hn in the original disparity map is defined as 0.

[0053] 4.2) Determining the Initial Road Profile: Disparity values ​​are characterized by being larger near and smaller far, and should be continuous and smooth. Points that do not conform to this characteristic are called singular points. In the V-disparity map, the disparity value corresponding to the point with the maximum intensity in each row is more likely to be the road surface than the other disparity values. Therefore, by removing the singular points in each row and updating the road surface deviation range for the singular point row, the initial road profile can be determined in the V-disparity map.

[0054] 4.3) Reclassification of Misclassified Points: Because singular points are judged based on the root disparity value, which is typically larger near and smaller far, and should be continuous and smooth, there may be bias. These misclassified points are called misclassified points, and therefore need to be reclassified. Misclassified points can cause road surface loss, so we first need to recover these lost points. Then, we use the minimum inter-class similarity algorithm to reclassify small areas of mislabeled points. Finally, we reassign values ​​to larger areas of continuous labeled misclassified points to correct the misclassified points.

[0055] 4.4) Determine the target detection area: In order to reduce the complexity of vehicle detection and simplify the detection process, the detection image is segmented based on the initial road contour, so that subsequent detection vehicles are only detected in the detection area.

[0056] 5) Perform pixel association on the target detection area determined in step 4) and the V disparity map from step 3), remove the disparity information of pixels below the road surface in the V disparity map, detect the vertical line where the obstacle intersects the road surface, and construct a new disparity map containing only the disparity information above the road surface. The process is as follows:

[0057] The oblique lines or curves in the V disparity map represent the initial contour of the road surface, the vertical line segments represent the position of the obstacle, the intersection is the obstacle's contact point, and the height of the vertical line segments represents the obstacle's height. Obstacles vary in shape, and the distances of various points on them from the acquisition device also vary. However, compared to other obstacles, the disparity values ​​on a particular obstacle are within a certain range and are nearly continuous. Therefore, under ideal conditions, the disparity statistical points corresponding to the obstacle in the V disparity map are also approximately a continuous vertical line segment, where the intersection of the vertical line segment and the straight line segment is the coordinate of the contact point between the vehicle and the road surface in the original vehicle image. A standard Hough transform is performed on the V disparity map to detect vertical lines, thereby detecting the vertical line where the obstacle intersects the road surface, that is, the vertical position information of the obstacle. The disparity information of pixels below the road surface in the V disparity map is eliminated through the target detection area determined in step 4).

[0058] 6) Based on the new disparity map obtained in step 5), a U disparity map is constructed and obstacle lines are detected. The process is as follows:

[0059] Since the U disparity map is to accumulate the number of pixels with the same disparity in each column in the U direction, the new disparity map is scanned column by column from top to bottom along the U direction, and the horizontal length of the disparity map is used as the horizontal coordinate of the U disparity map, and the accumulated number of pixels with the same disparity value in each column is used as the intensity of each coordinate point of the U disparity map, thereby forming a "disparity value-horizontal coordinate-number" map (i.e., U disparity map), as shown in the figure. Figure 11 shown.

[0060] Since the line segments in the U-disparity map represent the location of obstacles on the road, the horizontal width of the line segments represents the width of the obstacle. In real-world situations, obstacles vary in shape, and each point on them is at varying distances from the acquisition device. However, compared to other obstacles, the disparity values ​​for the same obstacle are nearly continuous within a certain range. The V-disparity map is used to generate a disparity map that eliminates the area below the road surface. The U-disparity map is then calculated for this new disparity map. A standard Hough transform is then applied to the U-disparity map to detect horizontal lines, i.e., obstacle line detection, to obtain the horizontal position of the obstacle.

[0061] 7) The straight lines detected in the V and U disparity maps are comprehensively verified on the left image of the road scene. Contour recognition is performed on pedestrians and vehicles among obstacles to distinguish them from non-standard obstacles. The UV disparity target detection results are then combined. The process and principle are as follows:

[0062] Since the projection of the ground in the V-disparity map is a diagonal line with increasing pixels per line, indicating a decreasing disparity, the position of the horizontal ground can be detected by detecting diagonal lines in the disparity map. Because the same obstacle is at almost the same distance from the camera, it appears as a vertical line in the V-disparity map. Therefore, calculating the length of the vertical line segment can be used to calculate the obstacle's vertical height. The position of the vertical line also represents the obstacle's distance from the camera. The closer the vertical line is to the right side of the V-disparity map, the closer it is to the camera. If the vertical line intersects the diagonal line, it proves that the obstacle is on the horizontal ground. The projection of the obstacle in the U-disparity map is a horizontal line. The length of the horizontal line represents the obstacle's width, and the position of the horizontal line in the U-disparity map also represents the distance from the camera. The closer the horizontal line is to the camera, the lower it appears in the U-disparity map projection. Therefore, combined with the UV-disparity map, the position of the obstacle and the horizontal ground relative to the camera can be determined. That is, the vertical position information of the obstacle obtained in step 5) and the horizontal position information of the obstacle obtained in step 6) are combined, that is, the detected obstacles are arranged according to the Y value in the U disparity map, and the detected obstacles are arranged according to the X value in the V disparity map. The two-dimensional coordinates of these obstacles are located in the original disparity map obtained in step 2), and the three-dimensional coordinates of these obstacles in the three-dimensional space are further calculated based on the disparity values ​​in the disparity map.

[0063] Among them, the contour recognition process of vehicles and pedestrians is as follows: using the stereoscopic vision characteristics of disparity, based on the principle that the disparity of points on the same target is similar, the location of the obstacle is preliminarily determined within the detection area of ​​the road surface. Since the distances of points on the same obstacle to the on-board camera are similar, that is, the depth values ​​are similar, and the parameters of the camera are all fixed values, the disparity values ​​of the points on the obstacle are also similar. Obstacles with similar vertical disparity values ​​can be obtained from the V disparity map, and obstacles with similar horizontal disparity values ​​can be obtained using the U disparity map. Thus, the area where obstacles with similar disparity values ​​are located can be determined. First, the disparity value of each pixel in the initial disparity map is traversed. In order to further identify vehicles and pedestrians, each disparity value is threshold filtered, and the disparity values ​​that do not meet the threshold are set to 0. Then all connected areas are marked, all connected areas are scanned one by one, and area threshold filtering is performed. Then, the empirical aspect ratio of vehicles and pedestrians is filtered, and finally the contour recognition results of pedestrians and vehicles are obtained to distinguish between pedestrians, vehicles and non-standard obstacles. The process is as follows Figure 12 shown.

[0064] 8) Input the UV disparity target detection result obtained in step 7) into the YOLOv5 model for joint training detection, output the detection result and determine the two-dimensional image coordinates and three-dimensional space coordinates in combination with the corresponding disparity image relationship. The process is as follows Figure 13 shown.

[0065] Based on the multi-scale depth information and the rough position of the target in the UV disparity target detection result obtained in step 7), the YOLOv5 deep learning model is used to obtain more accurate two-dimensional coordinates and two-dimensional bounding boxes to further identify the target. Because the disparity image output by the UV disparity map only has depth of field information and no other interference information, YOLOv5 can be used to achieve fast and accurate target recognition, and can avoid the interference of the background in the environment on the identified object, and can achieve sample-free target recognition. The algorithm obtained by combining the two can identify the required target more accurately and efficiently than the traditional target recognition algorithm. The specific steps are as follows:

[0066] 8.1) Use the UV disparity target detection result obtained in step 7) as the input of the entire YOLOV5 network.

[0067] 8.2) Perform multi-scale detection of pedestrians, vehicles, and non-standard obstacles in the target environment.

[0068] 8.3) Divide the input image into grids.

[0069] 8.4) For the divided grids, predict the obstacle borders.

[0070] 8.5) Remove windows with lower probability based on the threshold.

[0071] 8.6) Finally, the precise two-dimensional bounding box and target type of the final target are output, and the three-dimensional XYZ coordinate information of each obstacle is calculated based on the corresponding disparity values ​​in the original disparity map of step 2). The results of YOLOv5 combined with UV disparity detection in this embodiment are as follows: Figure 14 As shown, it can be seen that both the vehicles on the road and the road signs are detected.

Claims

1. The target recognition method based on UV disparity detection and YOLOv5 is characterized by: Here are the steps: 1) Collect the dedistorted left and right images of the road scene using a binocular camera; 2) Calculate the disparity based on the collected left and right images of the road scene to obtain the original disparity map; 3) constructing a V-disparity map based on the original disparity map obtained in step 2); 4) Divide the road surface according to the V-disparity map and determine the target detection area; 5) Pixel association is performed between the target detection area determined in step 4) and the V disparity map obtained in step 3), disparity information of pixels below the road surface in the V disparity map is removed, the vertical line where the obstacle intersects the road surface is detected, and a new disparity map containing only disparity information above the road surface is constructed; 6) Based on the new disparity map obtained in step 5), a U-disparity map is constructed and obstacle lines are detected; 7) The straight lines detected in steps 5) and 6) are comprehensively verified on the left image of the road scene. Contour recognition is performed on pedestrians and vehicles among obstacles to distinguish between pedestrians, vehicles and non-standard obstacles, and the UV parallax target detection results are combined; 8) Input the UV disparity target detection result obtained in step 7) into the YOLOv5 model for joint training detection, output the detection result and combine it with the corresponding original disparity map to determine the target's two-dimensional image coordinates and three-dimensional space coordinates.

2. The target recognition method based on UV parallax detection and YOLOv5 according to claim 1, characterized in that: The specific steps of step 4) are as follows: 4.1) Determining the Horizon: The road surface area is mapped in the V disparity map, with the lower left corner as the origin. The disparity in the V direction decreases as the coordinate increases. The disparity statistical point corresponding to the road surface area in the V disparity map is a continuous diagonal line segment, which represents the road surface area. This diagonal line segment contains the longitudinal extent of the road surface area in the original image. Because the road surface area extends from the bottom of the image to the horizon, the horizon position is determined first. First, the number of rows where the road surface contour stops extending to the left is set as the row where the horizon is located. The height Hn of the horizon row is determined from the input V disparity map. The position above the horizon is defined as the non-road area. The disparity of the rows above Hn in the original disparity map is defined as 0. 4.2) Determining the Initial Road Profile: Disparity values ​​are characterized by being larger near and smaller far away, and should be continuous and smooth. Points that do not conform to this characteristic are called singular points. The disparity value corresponding to the maximum intensity point in each row of the V-disparity map is more likely to be the road surface than the other disparity values. Therefore, the singular points in each row are removed, and the road surface deviation range for the singular point row is updated to determine the initial road profile in the V-disparity map. 4.3) Reclassification of Misclassified Points: Since singular points are judged based on the root disparity value, which is larger near and smaller far and should be continuous and smooth, there may be deviations. These deviated singular points are called misclassified points, so they must be reclassified. Misclassified points can cause road surface loss, so we first find these lost points. Then, we use the minimum inter-class similarity algorithm to reclassify small areas of mislabeled points. Finally, we reassign values ​​to misclassified points in larger continuous label areas to correct the misclassified points. 4.4) Determine the target detection area: Segment the detection image based on the initial road contour so that subsequent detection vehicles are only detected in the detection area.

3. The target recognition method based on UV parallax detection and YOLOv5 according to claim 1, characterized in that: The specific steps of step 5) are as follows: The oblique lines or curves in the V-disparity map represent the initial contour of the road surface, the vertical line segments represent the position of the obstacle, the intersection is the contact point of the obstacle, and the height of the vertical line segments represents the height of the obstacle; The disparity statistical point corresponding to the obstacle in the V disparity map is also approximately a continuous vertical line segment, where the intersection of the vertical line segment and the straight line segment is the coordinate of the contact point between the vehicle and the road surface in the original vehicle image; the V disparity map is subjected to a standard Hough transform to detect vertical lines, thereby detecting the vertical line where the obstacle intersects the road surface, that is, the vertical position information of the obstacle; and the disparity information of pixels below the road surface in the V disparity map is eliminated through the target detection area determined in step 4).

4. The target recognition method based on UV parallax detection and YOLOv5 according to claim 1, characterized in that: The specific steps of step 6) are as follows: The U disparity map is calculated for the new disparity map, and the horizontal line is screened by performing standard Hough transform detection on the U disparity map, that is, obstacle straight line detection is performed to obtain the horizontal position information of the obstacle.

5. The target recognition method based on UV parallax detection and YOLOv5 according to claim 1, characterized in that: The step 7) is specifically as follows: combining the vertical position information of the obstacle obtained in step 5) and the horizontal position information of the obstacle obtained in step 6), that is, arranging the detected obstacles according to the Y value in the U disparity map, arranging the detected obstacles according to the X value in the V disparity map, locating the two-dimensional coordinates of these obstacles in the original disparity map obtained in step 2), and further calculating the three-dimensional coordinates of these obstacles in the three-dimensional space through the disparity values ​​in the disparity map; wherein, the vehicle and pedestrian contour recognition method is as follows: first, traversing the disparity value of each pixel point in the initial disparity map, in order to further identify vehicles and pedestrians, performing threshold filtering on each disparity value, setting the disparity value that does not meet the threshold to 0, then marking all connected areas, scanning all connected areas one by one, performing area threshold filtering, and then filtering by the empirical aspect ratio of vehicles and pedestrians, and finally obtaining the contour recognition results of pedestrians and vehicles, thereby distinguishing pedestrians, vehicles and non-standard obstacles.

6. The target recognition method based on UV parallax detection and YOLOv5 according to claim 1, characterized in that: The specific steps of step 8) are as follows: 8.1) Use the UV disparity object detection results obtained in step 7) as the input of the entire YOLOV5 network; 8.2) Perform multi-scale detection of pedestrians, vehicles, and non-standard obstacles in the target environment; 8.3) Divide the input image into grids; 8.4) For the divided grid, predict the obstacle border; 8.5) Remove windows with low probability based on the threshold; 8.6) Finally, the precise 2D bounding box and target type of the final target are output, and the 3D XYZ coordinate information of each obstacle is calculated based on the corresponding disparity values ​​in the original disparity map of step 2).

Images