Fast Annotation And Detection Algorithm Based On Color Point Cloud Objects For Unmanned Driving

In recent years,with the demand for safety and free time,there has been a wave of unmanned research around the world.For unmanned driving,quickly and accurately sensing the surrounding environment of the vehicle is the key to the safe and stable of the unmanned vehicle.Among various sensors,Li DAR is used widely in the field of unmanned driving due to its high precision and unaffected by ambient light.The object detection algorithm based on point cloud data has made some progress,but it still has the following shortcomings in the field of unmanned driving:(1)high time complexity with long detection time;(2)low precision with too many missed targets or false targets;(3)difficulty in obtaining data labels used to train deep neural networks.Aiming at the problems faced by current point cloud object detection,This thesis designs a high-speed object detection algorithm for unmanned Li DAR point cloud based on three tasks: point cloud annotation,data preprocessing and network design.Point cloud annotation is mainly divided into three tasks: ground point removal,color point cloud synthesis and annotation tool design.Point cloud annotation is a two-step process: finding the target and framing the target with the appropriate three-dimensional frame.The presence of ground points connects most objects together,which not only increases the difficulty of finding the target makes it hard for three-dimensional frame to adjust the gap between the ground point and the target.In the original point cloud,the point only contains the object location and the information of Li DAR's reflection,and it is difficult to find the object.The color information is assigned to the color point cloud formed by the point cloud,making the texture information of object is more obvious,and it is more convenient to find the object.At the same time,a semi-automatic annotation tool is designed,after finding the object,to make the object frame size automatically fitted according to the object size.The preprocessing of data is mainly for the disorder and sparsity of point cloud data and the environmental information required for driverless driving.The object area is divided into point clouds,and the point cloud is processed into structured data which can be used in convolutional neural networks.In this thesis,a single-step object detection algorithm based on deep residual network structure is designed.The preprocessed data is directly input into the network,and features are extracted in the backbone network,and then object classification and attribute regression are performed in the header network.In this thesis,the accuracy and speed of the proposed algorithm are tested and quantified in the public dataset.The results show that under the condition of good detection accuracy,the detection speed of the algorithm is very fast,which meets the requirements of real-time performance,and achieves an ideal detection effect.