Pedestrian Detection And Recognition Using LiDAR

Driverless technologies have received more and more attention in the last decade.The sensing system is particularly important for unmanned vehicles,it can detect the surrounding environmental information and the vehicle status information,and provide the necessary information for control system.The unmanned vehicle often integrates multiple sensors to make unmanned vehicles more intelligent and safe while driving.Lidar is widely used in unmanned vehicle systems due to its large detection range,high ranging accuracy and freedom from illumination conditions.Target detection and pedestrian recognition based on three-dimensional LiDAR is divided into three parts: ground extraction,point cloud segmentation and pedestrian recognition.Ground extraction mainly includes establishing depth maps and removing ground point cloud data,converting point cloud data into a depth map and then removing ground point clouds.Point cloud segmentation is the extraction of different objects from non-ground point cloud data.Pedestrian recognition is performed by classifying the segmented target using a classifier.For the low real-time problem of traditional ground removal algorithm,this thesis proposes an angle threshold algorithm based on a depth map.With the mapping relationship between point cloud data and a depth map,the original data was transformed into a depth map.Removing the ground point cloud data by using the angle threshold of the LiDAR scanning line.For the segmentation failure of adjoining targets and real-time problem of segmentation algorithm,this thesis proposes an improved DBSCAN algorithm based on depth map.Considering the space Euclidean distance and angular distance when searching all neighbors in the Eps neighborhood The non-ground point cloud was clustered and segmented by the improved DBSCAN algorithm combined with the depth map and the adaptive parameter.Experimental results show that the proposed method has a significant improvement in time efficiency compared with the traditional clustering algorithm.Moreover,the under-segment error rate was decreased.In view of shortcomings of PointNet network in local feature description for the pedestrian recognition.In this thesis,the PointNet neural network model with multi-scale feature fusion is proposed,which combines multi-scale local features and global features to finally realize the recognition of pedestrian targets.The offline process mainly constructs the training sample library,and trains the PointNet neural network model with multi-scale feature fusion.The online process is mainly to screen the candidate pedestrian targets,and to convert the point cloud target to a fixed size,and then put it into the PointNet classification network for feature learning.The classifier determines whether the point cloud target is a pedestrian.Experimental results show that the model has a certain improvement in recognition accuracy compared to the PointNet network.