| With the development of 3D sensor technology,the application of 3D point cloud data in fields such as autonomous driving has become increasingly widespread.Thanks to the excellent spatial description capabilities of point cloud data,3D point cloud perception tasks,such as object detection and classification,and semantic segmentation,can extract critical information from massive and complex data,providing important data support for the overall functionality and operation of the system.Focusing on these tasks,this paper primarily investigates data augmentation techniques and the theoretical foundations and practical applications of self-supervised algorithms in 3D point cloud perception tasks,aiming to optimize and enhance the performance of 3D point cloud perception models.Under the funding of the National Natural Science Foundation of China Key Project(No.61934003),the main research work of this paper is as follows:Firstly,for LiDAR(Light Detection and Ranging)point cloud object detection algorithms,a data augmentation technique AdaMix(Adaptive Resampling of Multiscale Object Mixup AdaMix)is designed and implemented.By combining objectlevel point cloud mixing and part-level vehicle point cloud mixing algorithms,the diversity of 3D point cloud data is effectively enhanced.The object-level point cloud mixing algorithm enables free movement of object point clouds across frames in a scene.At the same time,the point cloud density is adaptively adjusted based on the object’s location,and it is given a line-scan style of LiDAR point cloud data,ensuring the authenticity and consistency of the point cloud data.The part-level vehicle point cloud mixing algorithm mixes different parts of multiple vehicle point clouds together to construct a complete shape vehicle point cloud,allowing arbitrary adjustment of the vehicle point cloud’s yaw angle.Additionally,a position sampling module and a scene point completion algorithm are proposed.The position sampling module ensures that the inserted object point clouds are placed in reasonable positions while maintaining the original distribution of object positions in the data.The scene point completion algorithm further enhances the realism of the generated point clouds.The AdaMix data augmentation technique is applied to three different 3D object detection networks:Point Pillars,PV-RCNN,and SECOND,and their performance is evaluated on the KITTI dataset,showing significant improvements in the accuracy of vehicle and pedestrian detection tasks.The specific results are as follows: for vehicle and pedestrian detection AP(moderate difficulty),on Point Pillars,they are improved by 3.50% and 11.35%,respectively;on PV-RCNN,they are improved by 0.16% and 14.35%,respectively;on SECOND,they are improved by 1.59% and 9.59%,respectively.These experimental evaluation results verify the effectiveness of the AdaMix data augmentation technique in improving the performance of LiDAR point cloud object detection.Secondly,by further adjusting and improving the AdaMix data augmentation technique,this algorithm is extended to LiDAR point cloud data semantic segmentation,proposing AdaMix-Seg.Specifically,the construction and extraction algorithm of the dense point cloud database in the object-level mixing algorithm is redesigned,and a parallel data processing strategy is introduced,significantly reducing the algorithm’s computation time,enabling real-time generation of new data during point cloud segmentation network training.At the same time,to address the limitations of the global mIoU metric in LiDAR point cloud segmentation evaluation,this metric is applied to local regions of LiDAR point clouds.Evaluation results on the Semantic KITTI dataset show that AdaMix-Seg increases the global mIoU metric of the Cylinder3 D network by 1.4%,and also achieves certain improvements in the local mIoU metric.Finally,to address the information loss issue in 3D point cloud shape classification due to the combination of point-wise multi-layer perceptrons and max pooling,a selfsupervised pre-training model based on autoencoders and a multi-grid autoencoder(MGA)is proposed.The pre-training model compresses and reconstructs the input 3D point clouds through MGA,enabling the encoder to effectively extract shape features of the point clouds,reducing information loss.By training the classifier separately and fine-tuning the entire classification network,the optimal performance of the entire network is achieved.Evaluation results show that this self-supervised pre-training model improves the OA metric of the Point Net++ network by 2.0% and 4.7% on the Model Net40 and Scan Object NN datasets,respectively.The experimental results of MGA demonstrate that this autoencoder can effectively extract valuable information from 3D point clouds in an unsupervised learning environment,and assist in enhancing the performance of classification networks.In summary,this paper is dedicated to exploring data augmentation techniques and selfsupervised learning algorithms in 3D point cloud perception tasks,aiming to partially enhance the performance of LiDAR point cloud data processing,especially in key tasks such as object detection,semantic segmentation,and shape classification.Through the evaluation and analysis of a series of experimental results,it is demonstrated that the above three algorithms have the potential to improve the diversity of 3D point cloud data and address information loss issues. |
