Research On Point Cloud Target Recognition For Autonomous Driving

Autonomous vehicles can sense the location of surrounding obstacles through sensors and identify category information.Based on this information,they can make decisions accordingly,thus driving safely and smoothly.Because of its high detection sensitivity,high range resolution,no distance blind area and strong anti-interference ability,Frequency Modulated Continuous Wave(FMCW)lidar has become the most potential high quality 3D point cloud image sensor for automatic driving system.However,due to the large amount of redundant 3D point cloud data obtained by FMCW lidar,the real-time performance of target recognition is poor,and the mutual occlusion between different targets makes the target recognition difficult.In view of the above problems,this paper takes the FMCW lidar point cloud data as the research object,and focuses on the simulation of the FMCW lidar ranging system,the generation of point cloud data,the segmentation of ground point cloud,the clustering of non-ground point cloud,and the research of target point cloud identification methods,which are as follows:1)The imaging mechanism of FMCW lidar was explored,the simulation imaging system of FMCW lidar was designed,and the sawtooth modulated light source model was established.According to the linear relationship between the frequency difference of transmitting signal and echo signal and the target distance,the influences of physical parameters such as transmitting signal bandwidth,sampling number and signal-to-noise ratio on ranging accuracy were found out: The larger the signal bandwidth,the more sampling points and the smaller the signal-to-noise ratio,the higher the ranging accuracy of the system.In this paper,the simulation imaging system was tested based on PC platform Simulink.The experimental results show that when the transmitting signal bandwidth is 1500 MHz and the sampling number is,the noise resistance of the system reaches-48 d B.When the signal-to-noise ratio is-40 d B,the ranging accuracy of the system reaches meters in the ranging range of 0-150 meters.In this paper,three-dimensional point cloud data is generated by combining the detection target distance with the beam emission Angle and using the point cloud coordinate solution formula,which lays a foundation for the subsequent research on target segmentation and recognition methods.2)Aiming at the problem that the input of ground point cloud reduces the speed and accuracy of target recognition,a segmentation method based on cloth simulation filtering and European-style clustering is proposed.Firstly,elastic and dynamic models are established to characterize the point cloud as the geometric shape of the cloth surface,and then the cloth that belongs to ground point cloud is removed to achieve ground point cloud segmentation.Secondly,the construction of KD-tree speeds up the nearest neighbor search speed,reduces the clustering segmentation time,solves the problem that the traditional European clustering segmentation algorithm is easy to cause the under-segmentation or over-segmentation of the target point cloud,and realizes the high-precision segmentation of the target point cloud.The experimental results show that the point cloud segmentation method in this paper can effectively reduce the redundant point cloud data input.The accuracy of target clustering segmentation is 92.9%,which is 13.57% higher than that of European clustering segmentation algorithm.The processing time of single frame is 0.174 s.3)Aiming at the problem that the recognition speed of target point cloud is slow due to the large amount of segmented target point cloud data,this paper undersamples the target point cloud through the apophase sampling method to reduce the input of point cloud data,thus speeding up the speed of data processing and target point cloud recognition.In view of the insufficient accuracy of Pointnet ++ network to identify blocked targets,a Res MLP module based on affine transformation was constructed.Based on the deep learning theory,a Res Point Net++ network model was established to explore the intelligent feature characterization mechanism of point cloud and realize the high-precision recognition of blocked targets.In this paper,the proposed method is trained and tested based on Model Net40 data set and KITTI data set.The experimental results show that on Model Net40 data set,the accuracy of the proposed network is 93.4% and the recognition speed is 386 frames/s,which is 2.7% higher than that of Point Net++ network,and the recognition speed is 25% higher.On KITTI data set,the average recognition accuracy of Res Point Net++ for targets with mild,moderate and severe occlusion is 80.08%,which is 23.13% higher than that of Point Net++ network.The ground point cloud segmentation,non-ground point cloud clustering and target recognition experiments of the urban road traffic scene simulated by FMCW Lidar were carried out by the proposed algorithm.The experimental results show that the accuracy of the proposed algorithm is 80.15% and the recognition time is 0.265 seconds for a frame of 200,000 point clouds.It meets the requirements of autonomous driving tasks,provides theoretical support for the engineering application of high-speed and accurate target recognition technology in the field of autonomous driving,and speeds up the pace of practical application of autonomous driving.