Object Detection And Recognition Of Autonomous Driving Based On Depth Information And Multiple Feature Vectors

The 3D point cloud data collected by 3D lidar has now become an important data support for the autonomous vehicle environment perception system.Its huge amount of data and rich environment information provide important data information for autonomous vehicle object detection and recognition,and also bring data processing challenges.In order to reduce the complexity of data processing,improve the accuracy of object detection and recognition,and ensure the safe driving of autonomous vehicles,this thesis uses VLP-16 and NVIDIA Jetson TX2 as the experimental hardware development platform,detect objects by an algorithm based on depth information,and extract multiple reliable feature vectors of the object to classify and recognize obstacles.Analyze the three-dimensional point cloud data captured by the lidar,convert the point cloud data from the original spherical coordinate system to the spatial rectangular coordinate system,construct a voxel map of the point cloud and perform road segmentation processing.Research on the three-dimensional object detection algorithm based on depth information to solve the problem of over and under segmentation of object clustering under different depth information.Using DST(Dempster-Shafer Theory)combined with multi-view fuzzy inference assignment method to realize three-dimensional object detection,reduce three-dimensional data to two-dimensional data,and greatly reduce the amount of data calculation.According to the uneven distribution of point cloud data caused by the lidar measurement mechanism,an improved object detection algorithm based on depth information is proposed.A mechanism that can adaptively select the size of structural elements based on depth information is added to solve the problem of over-and under-segmentation of object clustering under different depth information.On the premise of ensuring the real-time performance of the algorithm,improve the accuracy of the algorithm.Research on the three-dimensional obstacle classification and recognition algorithm is based on the BPNN(Back Propagation Neural Network)using multiple feature vectors to solve the problem of pedestrian and vehicle classification and recognition in the autonomous driving environment.Use the data collected by real vehicles in the campus to establish an experimental data set for obstacle recognition,and give a feature extraction scheme for threedimensional obstacles.Select multiple feature vector combinations as input data,and perform obstacle classification and recognition experiments based on BPNN and SVM respectively.The experimental results show that when the voxel map top view area,square box model volume,and three-dimensional voxel density are three feature vectors as input,BPNN has better accuracy and real-time performance in obstacle classification and recognition.Three representative and universal experimental road sections were selected on real campus roads for actual vehicle verification.The results of the improved dilation and corrosion algorithm based on depth information and the advantages of this algorithm are analyzed on real roads,and the performance of this algorithm is compared with the traditional dilation and corrosion algorithm in continuous frames.Experiments verify that this algorithm has better accuracy and real-time performance.The average running time of a single frame of 3D object detection is 125.47 ms,and the accuracy rate is 95.81%.Data analysis was performed on the classification and recognition effects of consecutive frames on three different road sections,and the experimental results and program running time of the three road sections were compared.The average running time of a single frame of the obstacle classification and recognition algorithm is 191.10 ms,and the accuracy is 93.26%.Experimental results show that the object detection and recognition algorithm based on depth information and multiple feature vectors has good accuracy and real-time performance.