| As a fundamental sensing device for autonomous driving,safety and defense,laser radar has been widely used in various industries.Compared with other 2D sensors,laser radar point cloud data has advantages such as smaller data size,faster processing speed,and more target dimension information,making it particularly suitable for use in pedestrian detection and tracking systems based on unmanned vehicles that require low visibility,high real-time performance,and high accuracy.To this end,this thesis conducts research on the laser radar pedestrian detection and tracking system based on unmanned vehicles in practical application scenarios.The specific contents are as follows:(1)This thesis studies the algorithm principles used in laser radar point cloud filtering and object segmentation,and conducts comparative experiments using public datasets and actual collected data to select filtering and segmentation algorithms suitable for "pedestrian detection and tracking".(2)This thesis proposes an improved DBSCAN clustering-based segmentation algorithm,which uses the geometric features of pedestrians to "reduce weight" some point clouds.Through experiments,it is shown that the improved DBSCAN algorithm can achieve a segmentation success rate of 86.4% when separating up to 5 targets,which is more than 5% higher than other segmentation algorithms and can effectively alleviate the problem of insufficient segmentation of pedestrian targets.(3)This thesis presents a multi-pedestrian trajectory tracking framework based on the segmented results of laser radar point clouds.In this framework,tracking filtering algorithms,association algorithms between targets and effective trajectories,and management of tracking trajectories are discussed.The effectiveness of the tracking framework and algorithm are demonstrated through simulation experiments and real point cloud data.The root mean square error of the target in the tracking framework can be reduced to less than 0.2m.(4)This thesis addresses the problem that traditional Point Net++ networks cannot combine bottom-level detail information with high-level semantic information and treat three-dimensional point cloud coordinates as input features.By adding higher-order moment features at the input end to expand the point cloud data dimension,and improving the network model by combining feature pyramid branch structures and using lateral connections to merge bottom-up and top-down feature maps,the improved Point Net++ network achieves an accuracy of 91.52% compared to traditional Point Net++ networks.(5)Compared with the traditional "detection followed by tracking" algorithm process,the "tracking followed by detection" processing method proposed in this paper has higher pedestrian trajectory accuracy,with a confidence level of up to 95.1% for single targets.In the "tracking followed by detection" processing framework,this thesis use the improved Point Net++ network to detect multi-frame point clouds,and eventually obtains target tracking trajectories with a trajectory confidence level of more than 78.6% for up to 5 targets. |
