Research On Data Augmentation And Real-time Target Recognition Algorithm For Point Cloud Data

With the continuous improvement of lidar sensor technology,the target recognition algorithm for point cloud data has gradually become a research hotspot.3D point cloud data has the advantages of being unaffected by weather,strong anti-interference ability,allweather operation,rich in information,and capable of ranging.However,point cloud data has the characteristics of sparseness and disorder,which makes the target recognition based on point cloud data having the following difficulties:(1)Due to the sparse characteristics of point cloud data,it only contains the outline information of the object,so the feature extraction is difficult.(2)Due to the disordered characteristics of point cloud data,it is difficult to directly identify objects in point cloud data.(3)Target recognition in road scenes requires the network to provide real-time results with limited computing power.In order to solve the above problems,this paper has carried out the following three researches:(1)Aiming at the problem that point cloud data is sparse and only contains object outline information,this paper proposes a target recognition method based on point cloud bird’s-eye view data enhancement.This method extends the image channel on the basis of the traditional bird’s-eye view,adds the features of point cloud density and reflection intensity,and enhances the internal features of the object,making the model more robust to target recognition with sparse point clouds in the distance.The accuracy of target recognition of the network is enhanced.(2)In view of the disordered characteristics of point clouds,this paper maps the 3D point cloud data into a 2D bird’s-eye view to avoid the disorder of the input data,and performs target recognition based on the bird’s-eye view.Aiming at the target recognition problem in the road scene,the anchor frame is optimized based on the clustering algorithm,and the Io U loss function is used to replace the traditional target frame coordinate loss,so that the network can obtain a better target recognition effect in the road scene.(3)In order to meet the real-time performance of target recognition in road scenes,this paper proposes a lightweight feature extraction network and applies it to the point cloud target recognition algorithm based on bird’s-eye view to balance the recognition speed and recognition accuracy of the network.The perception scheme ensures the real-time operation of the algorithm in the unmanned scene,and finally realizes the mapping of the recognition result from the target frame to the 3D point cloud,and builds an "end-to-end" network model for target recognition oriented to point cloud data.In the experiment,this paper uses the self-collected data set to conduct experiments on the proposed method,which proves that the feature enhancement method based on the internal feature of the point cloud can effectively enhance the feature information of the point cloud and improve the network recognition ability.The 3D target recognition algorithm based on the single-stage target recognition framework realizes the effective recognition of point cloud data in road scenes.The improved method finally achieves a m AP of 92.8%,which is8% higher than the baseline method.The feature extraction network is designed to be lightweight based on the depthwise separable convolution.Compared with the baseline network,the improved network parameters are reduced by 89.9%,and the calculation amount is reduced by 86.8%.The resulting network can run in real time on mobile or embedded devices while maintaining 89.6% m AP.