A road vehicle detection method based on feature layer fusion

Abstract

The present invention relates to the field of autonomous driving technology, specifically to a road vehicle detection method based on feature layer fusion, including: collecting basic original point cloud data features through laser radar, rasterizing point clouds, fitting parameters of ground models, calculating ground model parameters, acquiring images and image preprocessing, and detecting and identifying objects and judging ground models based on a YOLO detection network. The present invention improves on the algorithm based on raster map mapping, proposes a ground segmentation algorithm based on multiple regions, splits the ground point cloud into multiple regions for segmentation, effectively alleviates the phenomenon of under-segmentation caused by uneven road surface, slope, etc., obtains ground model parameters through calculation, and obtains ground model parameter structure by identifying objects based on YOLO, and accurately identifies the ground model by matching the corresponding ground model parameters with the opposite model structure.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a road vehicle detection method based on feature layer fusion. Background Art

[0002] The automatic driving system adopts advanced communication, computer, network and control technologies to achieve real-time and continuous control of the vehicle. It uses modern communication methods and directly faces the vehicle to realize two-way data communication between the vehicle and the ground. It has a fast transmission rate and a large amount of information. Subsequent tracking vehicles and control centers can know the exact location of the vehicle in front in a timely manner, making operation management more flexible, control more effective, and more adapted to the needs of vehicle automatic driving.

[0003] At present, in the field of autonomous driving, vehicles are required to identify objects in front of them on the road. However, the accuracy of overall object recognition by vehicles is not high, which seriously affects the safety of autonomous driving. Therefore, a road vehicle detection method based on feature layer fusion is proposed. Summary of the invention

[0004] The purpose of the present invention is to provide a road vehicle detection method based on feature layer fusion to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A road vehicle detection method based on feature layer fusion comprises the following steps:

[0007] Step S1, collecting basic original point cloud data features by laser radar: using laser radar to detect the three-dimensional ground, separating ground point cloud data from the point cloud data by using a gradient information obstacle detection method and a ground point cloud detection method using the depth information of the original point cloud data, wherein the ground point cloud data includes plane point cloud data and ground object point cloud data;

[0008] Step S2, point cloud rasterization: the separated ground point cloud data is regarded as a whole block, and then the minimum side length of the cuboid is set according to the actual size of the point cloud, and then the cuboid is divided into three-dimensional grids, that is, the point cloud is rasterized to form a multi-region segmentation of the point cloud data;

[0009] Step S3, fitting the parameters of the ground model:

[0010] (1) Randomly select three points in the same 3D grid point cloud after rasterization and calculate the normal vector of the plane where the three points are located by cross product of their vectors:

[0011] n=(P 2 -P 1 )X(P 3-P 1 )

[0012] Among them, P 1 =(x 1 ,y 1 , z 1 ), P 2 =(x 2 ,y 2 , z 2 ), P 3 =(x 3 ,y 3 , z 3 );

[0013] (2) Calculate the distance from any point in the point cloud to this plane:

[0014]

[0015] Among them, P i is any point in the point cloud, i=4,5,...,N;

[0016] (3) Setting the threshold d i <τ to extract normal point cloud data, save the point cloud that meets the conditions, form a point cloud data set, and record the number of point cloud data set points;

[0017] (4) Iterate steps (1) to (3) T times, and then save the point cloud data set with the largest number of points among all point cloud data sets;

[0018] (5) Steps (1) to (4) need to be repeated for each three-dimensional grid;

[0019] (6) After obtaining the point cloud data set saved by each three-dimensional grid, the saved point cloud data is fine-tuned using the least squares method, and the point cloud data on the model parameters are extracted therefrom;

[0020] (7) Iterate step (6) N times; since the fitting is random, set the ratio e of the outlier point cloud, where the outlier point cloud is the point cloud data on the non-model parameters; when the ratio e is set incorrectly, even within the maximum number of iterations N, the accurate ground point cloud is not extracted, and there is no need to perform N iterations, and set an expected normal value ratio E:

[0021] E=1-e

[0022] When the ratio e is set correctly and the number of point cloud data extracted on the model parameters is greater than the total number of point cloud data saved, the iteration is terminated; finally, the point cloud data extracted after fine-tuning is used to fit the accurate ground model parameters;

[0023] Step S4, calculate the ground model parameters: calculate the ground model parameters based on the point cloud data of the model parameters extracted from each grid point cloud

[0024]

[0025] in, is the wall normal vector, a, b, c correspond to the points in the X-axis, Y-axis, and Z-axis of the grid point cloud, respectively. A is the matrix composed of the extracted point cloud. A=[P 1 ...P S ] T , and S <N, is the constant term coefficient of the ground model,

[0026] Step S5, image acquisition and image preprocessing:

[0027] (1) Capturing images through a camera, and establishing a grid structure for the images through a GoogleNet model, dividing the grid structure to obtain small grids, wherein the size of the small grids is set to be the same in length, width, and height as the size of the three-dimensional grid obtained by point cloud segmentation in step S2;

[0028] (2) The input of the convolutional neural network is the image captured by the camera after the grid structure is divided in step (1). The convolutional neural network determines whether the center point of each small grid falls on the target, thereby deleting the non-target grids and retaining the grids with targets. The target parameters are predicted based on the retained grids. The predicted target parameters include the target category and the position of the target box.

[0029] (3) The target image obtained in step (2) is soft-sized normalized; secondly, convolutional neural network feature extraction is performed; the confidence of the bounding box is predicted; finally, the bounding box is filtered through the non-maximum suppression algorithm to obtain the ground model structure in the optimal image;

[0030] Step S6, detecting and identifying the object based on the YOLO detection network: by fusing the ground model parameters in step S4 with the ground model structure obtained in step S5, matching the corresponding ground model parameters with the ground model structure, outputting a fused target feature map, and outputting a target detection result;

[0031] Step S7, ground model determination: by comparing the target detection result output in step S6 with the database, the road condition ahead, whether the target obstacle is a vehicle and the type of vehicle are detected.

[0032] Furthermore, the obstacle detection method of the gradient information in step S1 is as follows: adjacent points are extracted from the neighboring scanning layer data, and two vectors are constructed, and then the gradient changes before and after the middle point are examined. A fixed ground point and obstacle point segmentation threshold is given to determine whether the middle point is a breakpoint. The above is used as a vertical interpretation of the original gradient information. Similarly, the same operation is performed in the horizontal data, and the ground point cloud data is separated from the point cloud data by traversing the horizontal and vertical data.

[0033] Furthermore, the depth information in step S1 is used for ground point cloud detection: the depth information of the original point cloud data is used to detect the ground point cloud based on the ground plane assumption, and the intervals between different layers of data, that is, the depth difference, are extracted from the original data, and compared with the layer data interval of the ideal plane to obtain ground point cloud data within a certain terrain range.

[0034] Furthermore, in step S2, the specific method of rasterizing the point cloud is:

[0035] A, calculation point set {P 1 , P 2 ..., Pi, ..., P N The maximum and minimum values ​​of the three coordinate axes XYZ:

[0036] X max =MAX(x 1 , x 2 , ..., x N ), X min =MIN(x 1 , x 2 , ..., x N )

[0037] Y max =MAX(y 1 ,y 2 , ..., y N ), Y min =MIN(y 1 ,y 2 , ..., y N )

[0038] Z max =MAX(z 1 , z 2 , ..., z N ), Z min =MIN(z 1 , z 2 , ..., z N )

[0039] Where Pi = [X i , Yi , Z i ] T , i = 1, 2, ..., N;

[0040] B. Determine the rasterization side length. The rasterization side length R determines the number of points in each grid and the calculation efficiency. The smaller the rasterization side length R, the more grids there are, the more computer resources are occupied, the lower the running speed, and the lower the efficiency. The larger the rasterization side length R, the lower the fitting stability of the ground point cloud, and the rasterization function is lost. Therefore, the rasterization side length R will be determined according to the experimental results. After determining the rasterization side length, the dimension of the point cloud grid can be calculated:

[0041]

[0042]

[0043]

[0044] C, calculate the index of each point after rasterization, encode the point cloud after rasterization, determine the number of the grid where each point is located, and the index h of each point in the grid:

[0045]

[0046]

[0047]

[0048] h=h x +h y *D x +h z *D x *D y

[0049] Among them, x, y, and z represent the X-axis, Y-axis, and Z-axis in the grid, respectively.

[0050] Furthermore, in the step S3, in which the parameters of the ground model are fitted, the appropriate number of iterations T in step (4) is derived as follows:

[0051]

[0052] Where, e: the ratio of abnormal points in point cloud data;

[0053] s: the number of points selected in each iteration;

[0054] T: maximum number of RANSAC iterations;

[0055] P: The probability of selecting a normal point at least once.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] This road vehicle detection method based on feature layer fusion is improved based on the algorithm of raster map mapping, and a ground segmentation algorithm based on multiple regions is proposed. The ground point cloud is split into multiple regions for segmentation, which effectively alleviates the under-segmentation phenomenon caused by uneven road surface and slope. The ground model parameters are obtained by calculation, and the ground model parameter structure is obtained by identifying the object based on YOLO. By matching the corresponding ground model parameters with the opposite model structure, the ground model can be accurately identified. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0059] Figure 2 A schematic diagram of obstacle detection using gradient information of the present invention;

[0060] Figure 3 It is a schematic diagram of ground point cloud detection based on depth information of the present invention. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0063] In the description of this patent, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", and "set" should be understood in a broad sense, for example, it can be fixedly connected or set, or it can be detachably connected or set, or connected or set in one piece. For ordinary technicians in this field, the specific meanings of the above terms in this patent can be understood according to specific circumstances.

[0064] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "several" is two or more, unless otherwise clearly and specifically defined.

[0065] Example 1

[0066] See also Figure 1 -and Figure 3 As shown, a technical solution provided by the present invention is:

[0067] A road vehicle detection method based on feature layer fusion, comprising:

[0068] First, the basic original point cloud data features are collected through laser radar:

[0069] Several laser radars are installed in front of the vehicle. By using the laser radar to detect the three-dimensional ground in front of the vehicle, obstacle detection based on gradient information can be used Figure 2 The schematic diagram is used to explain that adjacent points A, B, and C are extracted from the neighboring scan layer data, and two vectors AB and BC are constructed. Then, the gradient change before and after point B is examined. Given a fixed ground point and obstacle point segmentation threshold, it is determined whether B is a breakpoint. The above can be regarded as a vertical interpretation of the original gradient information. Similarly, similar operations can be performed in the horizontal data. The ground point cloud data can be separated from the point cloud data by traversing the horizontal and vertical data.

[0070] By treating the point cloud data as a whole block, the calculation is performed using the following formula:

[0071] A, calculation point set {P 1 , P 2 ..., Pi, ..., P N}(where Pi = [X i , Y i , Z i ] T , i=1,2,...,N) The maximum and minimum values ​​of the three coordinate axes XYZ:

[0072] X max =MAX(x 1 , x 2 , ..., x N ), X min =MIN(x 1 , x 2 ..., x N )

[0073] Y max =MAX(y 1 ,y 2 , ..., y N ), Y min =MIN(y 1 ,y 2 , ..., y N )

[0074] Z max =MAX(z 1 , z 2 , ..., z N ), Z min =MIN(z 1 , z 2 , ..., z N )

[0075] B. Determine the rasterization side length. The rasterization side length R determines the number of points in each grid and the calculation efficiency. The smaller the rasterization side length R, the more grids there are, the more computer resources are occupied, the lower the running speed, and the lower the efficiency. The larger the rasterization side length R, the lower the fitting stability of the ground point cloud, and the rasterization function is lost. Therefore, the rasterization side length R will be determined according to the experimental results. After determining the rasterization side length, the dimension of the point cloud grid can be calculated:

[0076]

[0077]

[0078]

[0079] C, calculate the index of each point after rasterization, encode the point cloud after rasterization, determine the number of the grid where each point is located, and the index h of each point in the grid:

[0080]

[0081]

[0082]

[0083] h=h x +h y *D x +h z *D x *D y

[0084] In the present invention, preferably, in the step S3 of fitting the parameters of the ground model, the derivation of the appropriate number of iterations T is:

[0085]

[0086] Where, e: the ratio of abnormal points in point cloud data;

[0087] s: the number of points selected in each iteration;

[0088] T: maximum number of RANSAC iterations;

[0089] P: The probability of selecting a normal point at least once

[0090] Then, the minimum side length of the cuboid is set according to the calculation structure of the point cloud, and the cuboid is divided into three-dimensional grids, that is, the point cloud is rasterized to form multi-region segmentation of the point cloud data, such as dividing the point cloud data into 9X9 three-dimensional grids to perform multi-region segmentation on the point cloud data;

[0091] Fit the parameters of the ground model:

[0092] (1) Randomly select three points in the same 3D grid point cloud after rasterization and calculate the normal vector of the plane where the three points are located by cross product of their vectors:

[0093] n=(P 2 -P 1 )X(P 3 -P 1 )

[0094] Among them, P 1 =(x 1 ,y 1 , z 1 ), P 2 =(x 2 ,y 2 , z 2 ), P 3 =(x 3 ,y 3 , z 3 );

[0095] (2) Calculate the distance from any point in the point cloud to this plane:

[0096]

[0097] Among them, P i is any point in the point cloud, i=4,5,...,N;

[0098] (3) Setting the threshold d i <τ to extract normal point cloud data, save the point cloud that meets the conditions, form a point cloud data set, and record the number of point cloud data set points;

[0099] (4) Iterate steps (1) to (3) T times, and then save the point cloud data set with the largest number of points among all point cloud data sets;

[0100] (5) Steps (1) to (4) need to be repeated for each three-dimensional grid;

[0101] (6) After obtaining the point cloud data set saved by each three-dimensional grid, the saved point cloud data is fine-tuned using the least squares method, and the point cloud data on the model parameters are extracted therefrom;

[0102] (7) Iterate step (6) N times; since the fitting is random, set the ratio e of the outlier point cloud, where the outlier point cloud is the point cloud data on the non-model parameters; when the ratio e is set incorrectly, even within the maximum number of iterations N, the accurate ground point cloud is not extracted, and there is no need to perform N iterations, and set an expected normal value ratio E:

[0103] E=1-e

[0104] When the ratio e is set correctly and the number of point cloud data extracted on the model parameters is greater than the total number of point cloud data saved, the iteration is terminated; finally, the point cloud data extracted after fine-tuning is used to fit the accurate ground model parameters;

[0105] Calculate the ground model parameters

[0106]

[0107] in, is the wall normal vector, A is the matrix composed of the extracted point cloud, A=[P 1 ...P S ] T , and S <N, is the constant term coefficient of the ground model,

[0108] Image acquisition and image preprocessing:

[0109] The image is collected through the camera, and each frame of the image taken by the camera is taken out, and a grid structure is established for the image through the GoogleNet model. At this time, the grid structure is the same as the grid structure in the point cloud data. Each frame of the picture is also gridded, and the picture is processed into a 9X9 grid that is the same as the point cloud data; such a grid output is also generated through the convolutional neural network. Each output in the grid predicts the target whose center point falls on this grid. The predicted target parameters include the category of the target and the position of the target box; the input image is soft-sized normalized; secondly, the convolutional network feature is extracted; the confidence of the bounding box is predicted; finally, the bounding box is filtered by the non-maximum suppression algorithm to obtain the ground model structure in the optimal picture; by fusing the ground model parameters with the ground model structure, the corresponding ground model parameters are matched with the opposite model structure, so that the length, width and height of the ground model structure can be clearly determined, and the fused target feature map is output, and the target detection result is output;

[0110] Finally, the ground model is judged by comparing the above output target detection results with the database to detect the road conditions ahead. Since the ground model structure has been combined with the ground model parameters, the specific data of the corresponding ground model, such as length, width and height, can be known. The corresponding model data can be compared with the data in the database to analyze the specific type of obstacles ahead, such as cars (sedans, vans or trucks).

[0111] Example 2

[0112] See also Figure 1 -and Figure 3 As shown, a technical solution provided by the present invention is:

[0113] A road vehicle detection method based on feature layer fusion, comprising:

[0114] First, the basic original point cloud data features are collected through laser radar:

[0115] Several laser radars are installed in front of the vehicle. The three-dimensional ground in front of the vehicle is detected by using the laser radar, and the ground point cloud detection method is used to detect the depth information. Figure 3 As shown, the detection of ground point cloud using the depth information of the original point cloud data is based on the ground plane assumption. The interval o between different layers of data, that is, the depth difference, is extracted from the original data, and compared with the layer data interval e of the ideal plane to obtain the ground point cloud data within a certain terrain range;

[0116] By treating the point cloud data as a whole block, as in Example 1, and then setting the minimum side length of the cuboid according to the actual size of the point cloud, and then dividing the cuboid into three-dimensional grids, that is, rasterizing the point cloud to form multi-region segmentation of the point cloud data, the point cloud data can be divided into 9X9 three-dimensional grids to perform multi-region segmentation on the point cloud data;

[0117] Fit the parameters of the ground model:

[0118] (1) Randomly select three points in the same 3D grid point cloud after rasterization and calculate the normal vector of the plane where the three points are located by cross product of their vectors:

[0119] n=(P 2 -P 1 )X(P 3 -P 1 )

[0120] Among them, P 1 =(x 1 ,y 1 , z 1 ), P 2 =(x 2 ,y 2 , z 2 ), P 3 =(x 3 ,y 3 , z 3 );

[0121] (2) Calculate the distance from any point in the point cloud to this plane:

[0122]

[0123] Among them, P i is any point in the point cloud, i=4,5,...,N;

[0124] (3) Setting the threshold d i <τ to extract normal point cloud data, save the point cloud that meets the conditions, form a point cloud data set, and record the number of point cloud data set points;

[0125] (4) Iterate steps (1) to (3) T times, and then save the point cloud data set with the largest number of points among all point cloud data sets;

[0126] (5) Steps (1) to (4) need to be repeated for each three-dimensional grid;

[0127] (6) After obtaining the point cloud data set saved by each three-dimensional grid, the saved point cloud data is fine-tuned using the least squares method, and the point cloud data on the model parameters are extracted therefrom;

[0128] (7) Iterate step (6) N times; since the fitting is random, set the ratio e of the outlier point cloud, where the outlier point cloud is the point cloud data on the non-model parameters; when the ratio e is set incorrectly, even within the maximum number of iterations N, the accurate ground point cloud is not extracted, and there is no need to perform N iterations, and set an expected normal value ratio E:

[0129] E=1-e

[0130] When the ratio e is set correctly and the number of point cloud data extracted on the model parameters is greater than the total number of point cloud data saved, the iteration is terminated; finally, the point cloud data extracted after fine-tuning is used to fit the accurate ground model parameters;

[0131] Calculate the ground model parameters

[0132]

[0133] in, is the wall normal vector, A is the matrix composed of the extracted point cloud, A=[P 1 ...P S ] T , and S <N, is the constant term coefficient of the ground model,

[0134] Image acquisition and image preprocessing:

[0135] The image is collected by the camera, and each frame of the image captured by the camera is taken out, and a grid structure is established for the image through the GoogleNet model. At this time, the grid structure is the same as the grid structure in the point cloud data. Each frame of the picture is also gridded, and the image is processed into a 9X9 grid that is the same as the point cloud data; such a grid output is also generated through the convolutional neural network, and each output in the grid predicts the target whose center point falls on this grid. The predicted target parameters include the category of the target and the position of the target box; the input image is soft-sized normalized; secondly, the convolutional network feature is extracted; the confidence of the bounding box is predicted; finally, the bounding box is filtered by the non-maximum suppression algorithm to obtain the ground model structure in the optimal picture; the object is detected and identified based on the YOLO detection network, and the ground model parameters in step S4 are fused with the ground model structure obtained in step S5, and the corresponding ground model parameters are matched with the opposite model structure, and the fused target feature map is output, and the target detection result is output;

[0136] Finally, the ground model is judged by comparing the above output target detection results with the database to detect the road conditions ahead. Since the ground model structure has been combined with the ground model parameters, the specific data of the corresponding ground model, such as length, width and height, can be known. The corresponding model data can be compared with the data in the database to analyze the specific type of obstacles ahead, such as cars (sedans, vans or trucks).

[0137] Based on the improvement of the raster map mapping algorithm, a multi-region ground segmentation algorithm is proposed. The ground point cloud is split into multiple regions for segmentation, which effectively alleviates the under-segmentation phenomenon caused by uneven road surface and slope. The ground model parameters are obtained by calculation, and the ground model parameter structure is obtained by identifying objects based on YOLO. By matching the corresponding ground model parameters with the opposite model structure, the ground model can be accurately identified.

[0138] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A road vehicle detection method based on feature-level fusion, characterized in that, it includes the following steps: Step S1, collect the features of the basic raw point cloud data through a lidar: Detect the three-dimensional ground by using a lidar, and separate the ground point cloud data from the point cloud data through an obstacle detection method based on gradient information or a ground point cloud detection method using the depth information of the raw point cloud data. The ground point cloud data includes plane point cloud data and object point cloud data on the ground; Step S2, point cloud rasterization: Regard the separated ground point cloud data as a whole block, then set the minimum side length of the cuboid according to the actual size of the point cloud, and then divide the cuboid into three-dimensional grids, that is, rasterize the point cloud to form a multi-region segmentation of the point cloud data; Step S3, fit the parameters of the ground model: (1) Randomly select three points in the point cloud of the same three-dimensional grid after rasterization and calculate the normal vector of the plane where the three points are located by the cross product of vectors; n=(P 2 -P 1 )X(P 3 -P 1 ) where, P 1 =(x 1 , y 1 , z 1 ), P 2 =(x 2 , y 2 , z 2 ), P 3 =(x 3 , y 3 , z 3 ); (2) Calculate the distance from any point in the point cloud to this plane; Among them, P i is any point in the point cloud, i=4,5,...,N; (3) Setting the threshold d i <τ to extract normal point cloud data, save the point cloud that meets the conditions, form a point cloud data set, and record the number of point cloud data set points; (4) Iterate steps (1) to (3) T times, and then save the point cloud data set with the largest number of points in all the point cloud data sets; (5) Each three-dimensional grid needs to repeat steps (1) to (4); (6) After obtaining the point cloud data set saved by each three-dimensional grid, use the least squares method to fine-tune the saved point cloud data and extract the point cloud data on the model parameters from it; (7) Iterate step (6) N times; Due to the randomness of fitting, set the ratio e of the outlier point cloud. The outlier point cloud is the point cloud data not on the model parameters; When the ratio e is set incorrectly, even within the maximum iteration number N, accurate ground point cloud is not extracted, then there is no need to perform N iterations, and a desired normal value ratio E is set: E = 1 - e When the ratio e is set correctly and the ratio of the point cloud data on the extracted model parameters to the total number of saved point cloud data is greater than E, terminate the iteration; Finally, use the fine-tuned extracted point cloud data to fit the accurate ground model parameters; In the step of fitting the parameters of the ground model in step S3, the derivation of the appropriate iteration number T in step (4) is: where, e: the ratio of outlier points in the point cloud data; s: the number of points selected each time; T: the maximum number of RANSAC iterations; P: the probability of selecting normal points at least once; Step S4, calculate the ground model parameters: calculate the ground model parameters based on the point cloud data of the model parameters extracted from each grid point cloud in, is the wall normal vector, a, b, c correspond to the points in the X-axis, Y-axis, and Z-axis of the grid point cloud, respectively. A is the matrix composed of the extracted point cloud. A=[P 1 ...P S ] T , and S <N, is the constant term coefficient of the ground model, Step S5, obtain the image and image preprocessing: (1) Collect the image through a camera, and establish a grid structure for the image through the GoogleNet model, divide the grid structure to obtain small grids, and set the length, width, and height of the small grids to be the same as the three-dimensional grid size of the point cloud segmentation in step S2; (2) The input of the convolutional neural network is the image collected by the camera after dividing the grid structure in step (1). Determine whether the center point of each small grid falls on the target through the convolutional neural network, delete the non-target grids accordingly, retain the grids with targets, and predict the target parameters through the retained grids. The predicted target parameters include the category of the target and the position of the target box; (3) The target image obtained in step (2) is soft-sized normalized; secondly, convolutional neural network feature extraction is performed; the confidence of the bounding box is predicted; finally, the bounding box is filtered through the non-maximum suppression algorithm to obtain the ground model structure in the optimal image; Step S6, detecting and identifying the object based on the YOLO detection network: by fusing the ground model parameters in step S4 with the ground model structure obtained in step S5, matching the corresponding ground model parameters with the ground model structure, outputting a fused target feature map, and outputting a target detection result; Step S7, ground model determination: by comparing the target detection result output in step S6 with the database, the road condition ahead, and whether the target obstacle is a vehicle and the type of vehicle are detected.

2. A road vehicle detection method based on feature layer fusion according to claim 1, Features: The obstacle detection method of the gradient information in step S1 is as follows: adjacent points are extracted from the neighboring scanning layer data, and two vectors are constructed. Then, the gradient changes before and after the middle points are examined. A fixed ground point and obstacle point segmentation threshold is given to determine whether the middle point is a breakpoint. The above is used as a vertical interpretation of the original gradient information. Similarly, the same operation is performed in the horizontal data. The ground point cloud data is separated from the point cloud data by traversing the horizontal and vertical data.

3. A road vehicle detection method based on feature layer fusion according to claim 1, Features: The method for detecting ground point cloud using depth information in step S1: Detecting ground point cloud using depth information of original point cloud data is based on the assumption of ground plane, extracting intervals between different layers of data, i.e., depth differences, from the original data, and comparing them with the intervals of layer data of an ideal plane to obtain ground point cloud data within a certain terrain range.

4. A road vehicle detection method based on feature layer fusion according to claim 1, Features: In step S2, the specific method of point cloud rasterization is: A, calculation point set {P 1 , P 2 ..., Pi, ..., P N The maximum and minimum values ​​of the three coordinate axes XYZ: X max =MAX(x 1 ,x 2 ,...,x N ),X mi n=MIN(x 1 ,x 2 ,...,x N ) AND max =MAX(and 1 ,and 2 ,...,and N ),AND min =MIN(and 1 ,and 2 ,...,and N ) WITH max =MAX(of 1 ,With 2 ,...,With N ),WITH min =MlN(from 1 ,With 2 ,...,With N ) Among them, Pi=[X i , Y i , Z i ] T , i = 1, 2, ..., N; B. Determine the rasterization side length. The rasterization side length R determines the number of points in each grid and the calculation efficiency. The smaller the rasterization side length R, the more grids there are, the more computer resources are occupied, the lower the running speed, and the lower the efficiency. The larger the rasterization side length R, the lower the fitting stability of the ground point cloud, and the rasterization function is lost. Therefore, the rasterization side length R will be determined according to the experimental results. After determining the rasterization side length, the dimension of the point cloud grid can be calculated: C, calculate the index of each point after rasterization, encode the point cloud after rasterization, determine the number of the grid where each point is located, and the index h of each point in the grid: h=h x +h y *D x +h z *D x *D y Among them, x, y, and z represent the X-axis, Y-axis, and Z-axis in the grid, respectively.

Images