Automatic driving data processing method and automatic driving traffic system

Abstract

The invention relates to the technical field of automatic driving, and particularly discloses an automatic driving data processing method and an automatic driving traffic system, and the method comprises the steps: 1, obtaining the obstacle information of a surrounding environment through the combination of a plurality of cameras and radars disposed on an automatic driving vehicle, carrying out the preprocessing of the collected original data, and carrying out the recognition of the collected data; pre-processed image data and radar point cloud information are obtained; 2, adopting a target detection algorithm based on deep learning, introducing a spatial pyramid network and an attention mechanism to detect obstacles in the image data, and obtaining an obstacle detection rectangular frame; according to the invention, obstacle information is obtained through combination of the camera and the radar, advanced technologies such as a deep learning algorithm, a spatial pyramid network, an attention mechanism and the like are utilized, the obstacle detection precision under complex road conditions is effectively improved, and meanwhile, radar point cloud is optimized by using methods such as a KD tree and principal component analysis and the like, so that the obstacle detection precision is improved. And the three-dimensional information of the obstacle is more accurate.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and in particular relates to an autonomous driving data processing method and an autonomous driving traffic system. Background Art

[0002] Nowadays, intelligence and safety have become necessary attributes of modern vehicles, so the autonomous driving technology of road vehicles has become the focus of technical research. The autonomous driving traffic system of road vehicles can make correct decisions by receiving and perceiving environmental information and then implement accurate and rapid control solutions, which can improve active safety and driving comfort.

[0003] With the rapid development of deep learning technology, target detection algorithms have also made great progress. However, in the field of autonomous driving, which has high real-time requirements, the current real-time target detection algorithm has low detection accuracy and cannot achieve satisfactory results. Target detection is one of the important links in the autonomous driving traffic system. Target detection collects and processes information through on-board sensors. The processed target information is used by the autonomous driving traffic system to make correct decisions and precise control. In order to ensure the comprehensiveness and accuracy of environmental information collection, autonomous driving vehicles currently generally use a combination of multiple cameras and radars to obtain target information, and perform data fusion processing on the signals collected by different types of sensors to achieve target detection. Because the camera is easily affected by factors such as occlusion, light, and road conditions. Therefore, target detection sometimes cannot accurately provide information about obstacles ahead. The detection distance of laser radar is long, and the perception information is three-dimensional information. However, the feature point cloud obtained is relatively sparse, which is difficult to meet the needs of the perception system.

[0004] Therefore, it is necessary to propose an autonomous driving data processing method and an autonomous driving traffic system to solve the problem of how autonomous driving vehicles can accurately and quickly detect obstacles ahead that exists in the prior art.

[0005] The above information disclosed in this background technology is only used to increase the understanding of the background technology of the present invention and therefore, it may include information that does not constitute the prior art known to ordinary technicians in this field. Summary of the invention

[0006] The purpose of the present invention is to provide an autonomous driving data processing method and an autonomous driving traffic system to solve the problems raised in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions: A method for processing autonomous driving data, comprising: Step 1: Obtain obstacle information of the surrounding environment through a combination of multiple cameras and radars deployed on the autonomous driving vehicle, pre-process the collected raw data, and obtain pre-processed image data and radar point cloud information; Step 2: Using a deep learning-based target detection algorithm, and introducing a spatial pyramid network and an attention mechanism to detect obstacles in the image data, and obtaining the obstacle detection rectangular frame; Step 3: Use KD tree to establish the topological relationship between the radar point cloud information, extract obstacle point cloud information, analyze the obstacle point cloud information by principal component analysis, and establish the circumscribed rectangle of the obstacle point cloud information based on the three-dimensional bounding box; Step 4: Projecting the circumscribed rectangle of the obstacle point cloud information onto the image data to form a corresponding point cloud target detection rectangular area, and fusing the point cloud target detection rectangular area with the obstacle detection rectangular frame to obtain actual obstacle information; Step 5: Use DeepSORT to associate the actual obstacle information with the actual obstacle information obtained in the previous detection, and use the motion model to predict the trajectory of the obstacle to obtain the predicted driving trajectory of the obstacle; Step six: infer whether there is a risk of collision between the driving path of the autonomous driving vehicle and the driving trajectory of the obstacle based on the predicted driving trajectory of the obstacle, and generate a prediction result.

[0008] Preferably, the preprocessing of the collected raw data includes: The raw data from multiple cameras and radars are synchronized by data fusion technology; the image data is grayed using a weighted average method to obtain grayed image data; image filtering is performed using the weight correlation between adjacent pixels as a weight to obtain filtered image data; At any time in the point cloud coordinate system The measured radar point cloud information is corrected and projected into the point cloud coordinate system Under the same coordinate system, the radar point cloud information is obtained; and the radar point cloud information is filtered by using a statistical filtering method to obtain the filtered radar point cloud information.

[0009] Preferably, after introducing the spatial pyramid network and attention mechanism, the algorithm framework of the deep learning-based target detection algorithm includes: A convolutional backbone network is used to extract features using a lightweight convolutional network as a backbone network; a spatial pyramid network is used to expand the feature extraction level using the spatial pyramid network to capture obstacle information at different scales; an attention mechanism is used to introduce an attention mechanism to adjust the degree of attention of the convolutional backbone network and the spatial pyramid network to the image data; in the feature extraction stage, a high-resolution local attention mechanism is introduced to focus on the area where the obstacle is located, and in the candidate area generation stage, an adaptive attention mechanism is used to adjust the importance of each candidate area; a multi-layer perceptron is used for target category classification and target bounding box regression prediction.

[0010] Preferably, before using the multi-layer perceptron for classification and prediction, the features of the convolutional backbone network are fused with the features of the spatial pyramid network to construct a complete contextual semantic representation; when using the multi-layer perceptron for classification and prediction, the above-mentioned fused features are used as input, and the fully connected layer is used to perform the target category classification and target bounding box regression prediction respectively; the binary cross entropy loss function is used to optimize the target category classification process, and the L1 and GloU loss functions are used as regression loss functions to optimize the target bounding box regression prediction process; non-maximum suppression is used to reduce repetitive detection between overlapping targets.

[0011] Preferably, the road point cloud is separated according to the actual road surface information using a point cloud segmentation algorithm for the ground to obtain the radar point cloud information of the separated road point cloud; a Euclidean distance k-neighborhood fast search algorithm is used to search based on the radar point cloud information of the separated road point cloud to extract a single obstacle point cloud information; a single obstacle point cloud information is analyzed by principal component analysis to obtain a surface normal vector of the obstacle point cloud information, and an OBB bounding box of the obstacle, that is, a circumscribed rectangle of the obstacle point cloud information, is obtained by calculation.

[0012] Preferably, the acquisition frequencies of the multiple cameras and radars are synchronized based on the radar acquisition cycle, and the data acquisition time of the multiple cameras and radars is synchronized to obtain the synchronized multiple cameras and radars; the world coordinate system is used as the bridge joint point cloud coordinate system The relationship between the pixel coordinate system W and the point cloud coordinate system Match the pixel coordinate system W to obtain the pixel coordinate system W and point cloud coordinate system after coordinate matching ; Use Kalman filtering to fuse the point cloud target detection rectangular area with the obstacle detection rectangular frame to obtain actual obstacle information.

[0013] Preferably, the appearance feature distance is obtained by calculating the cosine similarity between the feature vectors of the actual obstacle information and the actual obstacle information obtained in the previous detection, and the similarity of the target appearance is measured by the appearance feature distance; the distance between the actual obstacle information and the actual obstacle information obtained in the previous detection is measured by comparing the Euclidean distance of the center point or position of the bounding box; and each actual obstacle information is associated with an actual obstacle information obtained in the previous detection, and its state is updated through a Kalman filter.

[0014] Preferably, the motion model is selected based on the actual obstacle information, and a motion prediction equation of the actual obstacle information is established according to the selected motion model; the motion trajectory of the obstacle within a preset time period is calculated according to the actual obstacle information and the motion prediction equation; and a Kalman filter is used to fuse the above motion trajectory with the actual obstacle information actually detected to obtain a fused actual obstacle motion trajectory.

[0015] Preferably, the calculation predicts the intersection of the driving trajectory of the obstacle and the driving path of the autonomous driving vehicle within a preset time period; if there is an intersection or the two enter the same spatial area within the same time window, it is considered that there is a collision risk.

[0016] An autonomous driving traffic system, comprising: A memory including at least one set of instructions configured to implement a driving strategy for an autonomous vehicle; a processor, in a working state, reading the at least one set of instructions in the memory and generating corresponding decision information according to the at least one set of instructions; An execution instruction is issued to the autonomous driving vehicle according to the decision information.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The present invention obtains obstacle information through a combination of camera and radar, and uses deep learning algorithms, spatial pyramid networks, attention mechanisms and other advanced technologies to effectively improve the obstacle detection accuracy under complex road conditions. At the same time, the radar point cloud is optimized using methods such as KD tree and principal component analysis, making the three-dimensional information of the obstacle more accurate, thereby improving the comprehensive perception of static and dynamic obstacles. Through the fusion of image data and point cloud information, not only the positioning accuracy of obstacles is enhanced, but also the problem of incomplete information or misjudgment that may exist in a single sensor is solved. Through the DeepSORT algorithm for target association and motion trajectory prediction, it is possible to effectively track the dynamic changes of obstacles, deduce their future motion trajectories in real time, and conduct collision risk assessment between the driving path of the autonomous driving vehicle and obstacles.

[0018] Through the synergy of multiple perception, deep learning and data fusion technologies, the present invention not only improves the accuracy and real-time performance of obstacle detection, but also provides more reliable safety protection for autonomous driving traffic systems, helping vehicles make fast and accurate decisions in complex environments.

[0019] The above summary is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments and features described above, further aspects, embodiments and features of the present invention will be readily apparent by reference to the accompanying drawings and the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of the autonomous driving data processing method of the present invention. DETAILED DESCRIPTION

[0021] 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.

[0022] Embodiment 1:

[0023] See also Figure 1 As shown, a method for processing autonomous driving data includes: Step 1: Obtain obstacle information of the surrounding environment through a combination of multiple cameras and radars deployed on the autonomous driving vehicle, pre-process the collected raw data, and obtain pre-processed image data and radar point cloud information; Step 2: Use a deep learning-based target detection algorithm, introduce a spatial pyramid network and an attention mechanism to detect obstacles in the image data, and obtain an obstacle detection rectangle. Step 3: Use KD tree to establish the topological relationship between radar point cloud information, extract obstacle point cloud information, analyze the obstacle point cloud information through principal component analysis, and establish the obstacle point cloud information circumscribed rectangle based on the three-dimensional bounding box; Step 4: Project the circumscribed rectangle of the obstacle point cloud information onto the image data to form a corresponding point cloud target detection rectangular area, and merge the point cloud target detection rectangular area with the obstacle detection rectangular frame to obtain the actual obstacle information; Step 5: Use DeepSORT to associate the actual obstacle information with the actual obstacle information obtained from the previous detection, use the motion model to predict the trajectory of the obstacle, and obtain the predicted obstacle trajectory; Step 6: Based on the predicted driving trajectory of the obstacle, it is inferred whether there is a risk of collision between the driving path of the autonomous driving vehicle and the driving trajectory of the obstacle, and a prediction result is generated.

[0024] Synchronize raw data from multiple cameras and radar combinations through data fusion technology to ensure temporal consistency; The weighted average method is used to grayscale the image data captured by the camera. The formula is as follows:

[0025] In the formula, , and The coordinates are The R, G, and B component values ​​of the pixel point, , , Represents the weights of R, G, and B respectively. According to the different sensitivity of human eyes to R, G, and B, , , is the weight coefficient of R, G, and B.

[0026] The values ​​of R, G, B ( , , ) is usually based on the sensitivity of the human eye to light of different wavelengths. The following are the key materials and theoretical basis for support: 1. Human Eye’s Color Sensitivity The human eye has different sensitivities to different colors of light. Through experimental research, it is found that the human eye is most sensitive to green, followed by red, and finally blue. This difference in sensitivity is due to the distribution and response characteristics of the three types of cone cells in the retina. Among them: Red cones are most sensitive to long wavelengths of light (about 564-580nm); Green cones are most sensitive to light of medium wavelengths (about 534-565nm); Blue cones are most sensitive to short wavelengths of light (about 420-440nm).

[0027] These differences in sensitivity are quantified based on the natural perception of color by the human eye, resulting in different weighting factors for each color.

[0028] 2. Historical Basis of Weighting Coefficients These weight coefficients ( , , ) comes from television standards and color space models, especially the calculation method of the brightness (Y) component in the YUV color model. In the YUV color model, the brightness signal Y is obtained by weighting the RGB signals. The Y signal represents the brightness information of the image, so different weight coefficients are used to reflect the sensitivity of the human eye to color brightness.

[0029] Similar weighting factors are used in the ITU-R BT.601 standard for standard-defined video compression and coding:

[0030] This formula has been experimentally proven to show that green contributes the most to the brightness perceived by the human eye, followed by red, and blue contributes the least. These coefficients are widely used in television, video, and image processing to ensure that the visual effect conforms to the perception habits of the human eye.

[0031] 3. Color Science and Psychophysics From a psychophysical point of view, the perception of color depends not only on the wavelength of light, but also on the way the brain processes it. Scientists experimentally measure the intensity of the human eye's response to light stimuli of different colors and use this data to define color weighting coefficients. These weighting coefficients are derived from a large number of perception experiments and data.

[0032] The weight correlation between adjacent pixels is used as the weight for image filtering; Will At any time in the point cloud coordinate system The measured radar point cloud information is corrected and projected into the point cloud coordinate system Under the same coordinate system, the radar point cloud information is obtained to reduce the influence of motion distortion; The radar point cloud information is filtered using the statistical filtering method to obtain the filtered radar point cloud information.

[0033] After introducing the spatial pyramid network and attention mechanism, the target detection algorithm based on deep learning has the following algorithm framework: Convolutional backbone network, used for feature extraction using a lightweight convolutional network as the backbone network; Spatial pyramid network, which is used to expand the feature extraction level using the spatial pyramid network to capture obstacle information of different scales; Attention mechanism, which is used to introduce attention mechanism to adjust the degree of attention paid by convolution backbone network and spatial pyramid network to image data; Multilayer perceptron, used for classification of target categories and prediction of target bounding box regression.

[0034] Before using the multi-layer perceptron for classification and prediction, the features of the convolutional backbone network are fused with the features of the spatial pyramid network to construct a complete contextual semantic representation; When using the multi-layer perceptron for classification and prediction, the above fused features are used as input, and the fully connected layer is used for target category classification and target bounding box regression prediction respectively; The binary cross entropy loss function is used to optimize the target category classification process, and the L1 and GloU loss functions are used as regression loss functions to optimize the prediction process of target bounding box regression; non-maximum suppression is used to reduce repeated detections between overlapping targets.

[0035] The road surface point cloud is separated using a ground point cloud segmentation algorithm according to the actual road surface information to obtain radar point cloud information of the separated road surface point cloud; Use the Euclidean distance k-neighborhood fast search algorithm to search based on the radar point cloud information of the separated road point cloud to extract the single obstacle point cloud information; The KD tree is a European point cloud data structure suitable for k-dimensional space. First, a two-dimensional KD tree is constructed in a conventional way with a set of two-dimensional plane points, and then the principle is extended to three dimensions: (1) First, construct the root node, sort the set of two-dimensional plane points from small to large, and calculate the median of these points; (2) Sort the mount subtree point set based on the median value; (3) Construct the root node of the left subtree and use this root node as the basis to sort and mount the subtree point set; (4) Construct the right subtree root node and use this root node as the basis to sort and mount the subtree point set.

[0036] The process of the fast search algorithm based on Euclidean distance k-neighborhood is as follows: ① Establish a KD tree and define the pointer list L of the point cloud to be processed and the clustering result list S; ② Import all point cloud data P in L and find a certain point Put it into list S, use KD tree search to find its k neighbors, and calculate the distance d from these k points to list S respectively; ③ Calculate the adaptive threshold r through the coordinates of the center point of list S and the point cloud in L, and put the points whose d is less than r into list S; ④Find a point in list S except Other points , repeat step ②, and repeat this step until all points that meet the conditions are put into list S.

[0037] The point cloud information of a single obstacle is analyzed by principal component analysis to obtain the surface normal vector of the obstacle point cloud information. The OBB bounding box of the obstacle, that is, the circumscribed rectangle of the obstacle point cloud information, is obtained by calculation.

[0038] A single obstacle clustering set is a set of three-dimensional point clouds. The surface normal vector of the bounding box of the clustering target, that is, the principal axis of the OBB bounding box, can be obtained through principal component analysis. The projection direction of the point with the largest variance is taken as the main component. The first point with the largest variance is removed and the variance is calculated again. The projection direction of the point with the largest variance is taken as the second main component. The following component order can be obtained in the same way. The number of the final main components is less than or equal to the original space dimension C.

[0039] The acquisition frequencies of multiple cameras and radars are synchronized based on the radar acquisition cycle, and the data acquisition time of multiple cameras and radars is synchronized to obtain multiple synchronized cameras and radars; Use the world coordinate system to establish the point cloud coordinate system for the bridge The relationship between the pixel coordinate system W and the point cloud coordinate system Match the pixel coordinate system W to obtain the pixel coordinate system W and point cloud coordinate system after coordinate matching ; Kalman filtering is used to fuse the point cloud target detection rectangular area with the obstacle detection rectangular frame to obtain the actual obstacle information.

[0040] Embodiment 2:

[0041] See also Figure 1 As shown, this embodiment is basically the same as the above embodiment, except that the appearance feature distance is obtained by calculating the cosine similarity between the actual obstacle information and the feature vector of the actual obstacle information obtained by the previous detection, and the appearance feature distance is used to measure the similarity of the target appearance; By comparing the Euclidean distance of the center point or position of the bounding box, the distance between the actual obstacle information and the actual obstacle information obtained by the previous detection is measured; Each actual obstacle information is associated with the actual obstacle information obtained in the previous detection, and its status is updated through the Kalman filter.

[0042] Selecting a motion model based on actual obstacle information, and establishing a motion prediction equation of the actual obstacle information according to the selected motion model; Calculate the movement trajectory of the obstacle within a preset time period based on the actual obstacle information and the motion prediction equation; The Kalman filter is used to fuse the above motion trajectory with the actual obstacle information actually detected to obtain a fused actual obstacle motion trajectory.

[0043] Based on the trajectory of the obstacle and the path of the autonomous vehicle, determine the relative position of the two; Calculate the intersection of the predicted obstacle's driving trajectory and the autonomous vehicle's driving path within a preset time period; If there is an intersection or the two vehicles enter the same spatial area within the same time window, it is considered that there is a risk of collision.

[0044] As can be seen from the above, the present invention obtains obstacle information through a combination of camera and radar, and uses advanced technologies such as deep learning algorithms, spatial pyramid networks, and attention mechanisms to effectively improve the obstacle detection accuracy under complex road conditions. At the same time, the radar point cloud is optimized using methods such as KD tree and principal component analysis, making the three-dimensional information of the obstacle more accurate, thereby improving the comprehensive perception of static and dynamic obstacles. Through the fusion of image data and point cloud information, not only the positioning accuracy of the obstacle is enhanced, but also the problem of incomplete information or misjudgment that may exist in a single sensor is solved. Through the DeepSORT algorithm for target association and motion trajectory prediction, it is possible to effectively track the dynamic changes of obstacles, deduce their future motion trajectories in real time, and conduct collision risk assessments on the driving path of the autonomous driving vehicle and obstacles.

[0045] Embodiment 3:

[0046] An embodiment of the present invention also provides an autonomous driving traffic system, comprising: a memory, the memory comprising at least one set of instructions, the instructions being constructed to complete the driving strategy of an autonomous driving vehicle; a processor, reading at least one set of instructions from the memory in a working state, and generating corresponding decision information based on at least one set of instructions; and issuing execution instructions to the autonomous driving vehicle based on the decision information.

[0047] The embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, each process of the above-mentioned automatic driving traffic system embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. Among them, the computer-readable storage medium, such as read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random ACGess Memory, referred to as RAM), disk or optical disk, etc.

[0048] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0049] In the drawings of the embodiments disclosed in the present invention, only the structures involved in the embodiments disclosed in the present invention are involved, and other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of the present invention can be combined with each other.

[0050] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0051] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for processing autonomous driving data, characterized in that: include: Step 1: Obtain obstacle information of the surrounding environment through a combination of multiple cameras and radars deployed on the autonomous driving vehicle, pre-process the collected raw data, and obtain pre-processed image data and radar point cloud information; Step 2: Using a deep learning-based target detection algorithm, and introducing a spatial pyramid network and an attention mechanism to detect obstacles in the image data, and obtaining the obstacle detection rectangular frame; Step 3: Use KD tree to establish the topological relationship between the radar point cloud information, extract obstacle point cloud information, analyze the obstacle point cloud information by principal component analysis, and establish the circumscribed rectangle of the obstacle point cloud information based on the three-dimensional bounding box; Step 4: Projecting the circumscribed rectangle of the obstacle point cloud information onto the image data to form a corresponding point cloud target detection rectangular area, and fusing the point cloud target detection rectangular area with the obstacle detection rectangular frame to obtain actual obstacle information; Step 5: Use DeepSORT to associate the actual obstacle information with the actual obstacle information obtained in the previous detection, and use the motion model to predict the trajectory of the obstacle to obtain the predicted driving trajectory of the obstacle; Step six: infer whether there is a risk of collision between the driving path of the autonomous driving vehicle and the driving trajectory of the obstacle based on the predicted driving trajectory of the obstacle, and generate a prediction result.

2. The method for processing automatic driving data according to claim 1, characterized in that: The preprocessing of the collected raw data includes: Synchronize raw data from multiple cameras and radar combinations through data fusion technology; Performing image grayscale processing on the image data using a weighted average method to obtain grayscale image data; Using the weight correlation between adjacent pixels as the weight to perform image filtering processing to obtain the filtered image data; Will At any time in the point cloud coordinate system The measured radar point cloud information is corrected and projected into the point cloud coordinate system Under the above conditions, the radar point cloud information in the same coordinate system is obtained; The radar point cloud information is filtered using a statistical filtering method to obtain the filtered radar point cloud information.

3. The method for processing automatic driving data according to claim 2, characterized in that: After introducing the spatial pyramid network and attention mechanism, the algorithm framework of the deep learning-based object detection algorithm includes: Convolutional backbone network, used for feature extraction using a lightweight convolutional network as the backbone network; A spatial pyramid network is used to expand the feature extraction level using a spatial pyramid network to capture obstacle information at different scales; An attention mechanism is used to introduce an attention mechanism to adjust the degree of attention paid by the convolution backbone network and the spatial pyramid network to the image data; Multilayer perceptron, used for classification of target categories and prediction of target bounding box regression.

4. The method for processing automatic driving data according to claim 3, characterized in that: When the multi-layer perceptron is used to classify the target category and predict the target bounding box regression, it includes: Before using the multi-layer perceptron for classification and prediction, the features of the convolutional backbone network are fused with the features of the spatial pyramid network to construct a complete contextual semantic representation; When the multi-layer perceptron is used for classification and prediction, the fused features are used as input, and the fully connected layer is used to perform the target category classification and target bounding box regression prediction respectively; The binary cross entropy loss function is used to optimize the target category classification process, and the L1 and GloU loss functions are used as regression loss functions to optimize the target bounding box regression prediction process.

5. The method for processing automatic driving data according to claim 4, characterized in that: The method of using a KD tree to establish a topological relationship between the radar point cloud information, extracting obstacle point cloud information, analyzing the obstacle point cloud information by a principal component analysis method, and establishing a circumscribed rectangle of the obstacle point cloud information based on a three-dimensional bounding box includes: Separating the road surface point cloud using a point cloud segmentation algorithm for the ground according to the actual road surface information to obtain the radar point cloud information of the separated road surface point cloud; Using a fast search algorithm based on Euclidean distance k-neighborhood to search based on the radar point cloud information separated from the road point cloud, extracting a single obstacle point cloud information; The single obstacle point cloud information is analyzed by principal component analysis to obtain the surface normal vector of the obstacle point cloud information, and the OBB bounding box of the obstacle, that is, the circumscribed rectangle of the obstacle point cloud information, is obtained through calculation.

6. The method for processing automatic driving data according to claim 5, characterized in that: The projecting the circumscribed rectangle of the obstacle point cloud information onto the image data to form a corresponding point cloud target detection rectangular area, and fusing the point cloud target detection rectangular area with the obstacle detection rectangular frame to obtain actual obstacle information, including: Synchronize the acquisition frequencies of the multiple cameras and radars based on the radar acquisition cycle, synchronize the data acquisition time of the multiple cameras and radars, and obtain the synchronized multiple cameras and radars; Use the world coordinate system to establish the point cloud coordinate system for the bridge The relationship between the pixel coordinate system W and the point cloud coordinate system Match the pixel coordinate system W to obtain the pixel coordinate system W and point cloud coordinate system after coordinate matching ; The point cloud target detection rectangular area is merged with the obstacle detection rectangular frame using Kalman filtering to obtain actual obstacle information.

7. The method for processing automatic driving data according to claim 6, characterized in that: The using DeepSORT to associate the actual obstacle information with the actual obstacle information obtained from a previous detection includes: By calculating the cosine similarity between the actual obstacle information and the feature vector of the actual obstacle information obtained in the previous detection, an appearance feature distance is obtained, and the similarity of the target appearance is measured by the appearance feature distance; Measuring the distance between the actual obstacle information and the actual obstacle information obtained in a previous detection by comparing the Euclidean distance of the center point or position of the bounding box; Each piece of actual obstacle information is associated with the actual obstacle information obtained in a previous detection, and its state is updated through a Kalman filter.

8. The method for processing automatic driving data according to claim 7, characterized in that: The using the motion model to predict the trajectory of the obstacle includes: Selecting a motion model based on the actual obstacle information, and establishing a motion prediction equation for the actual obstacle information according to the selected motion model; Calculating the movement trajectory of the obstacle within a preset time period according to the actual obstacle information and the motion prediction equation; The Kalman filter is used to fuse the above motion trajectory with the actual obstacle information actually detected to obtain a fused actual obstacle motion trajectory.

9. The method for processing automatic driving data according to claim 8, characterized in that: The inferring whether there is a risk of collision between the driving path of the autonomous driving vehicle and the driving trajectory of the obstacle according to the predicted driving trajectory of the obstacle includes: Calculating the intersection of the predicted driving trajectory of the obstacle and the driving path of the autonomous driving vehicle within a preset time period; If there is an intersection or the two vehicles enter the same spatial area within the same time window, it is considered that there is a risk of collision.

10. An automatic driving traffic system, characterized in that: include: A memory including at least one set of instructions configured to implement a driving strategy for an autonomous vehicle; a processor, in a working state, reading the at least one set of instructions in the memory and generating corresponding decision information according to the at least one set of instructions; An execution instruction is issued to the autonomous driving vehicle according to the decision information.

Images