Research On Deep Learning-based Object Detection Method In Point Cloud

The research of image-based object detection has a long history,and it has many extensive applications now,such as face recognition,license plate detection and invoice detection.The main feature of these application scenarios is that they do not require accurate depth information in the target regions,and they pay more attention to the two-dimensional position and content of the target regions.However,in order to complete basic functions such as vehicle path planning,behavior control,and obstacle avoidance in autonomous driving scenarios,we need to obtain accurate location information of obstacles such as surrounding vehicles and pedestrians.In addition,if depth cameras are used to infer the depth information of targets in the environment,the accuracy of the results is low,which may lead to accidents such as vehicle collisions.Lidar sensors are more suitable for autonomous driving scenarios due to their wide detection range and stable data accuracy.Therefore,the research of object detection algorithm based on point cloud is of great significance.Early object detection algorithms based on point cloud mainly used machine learning methods,but such methods have poor versatility and long calculation time,and not satisfy demands for complex scenes and high real-time.Therefore,our article mainly studies the point cloud object detection algorithm based on deep learning,and attempts to improve the feature extraction ability and detection accuracy of the point cloud object detection algorithm in limited data.First,our article introduces point cloud data and datasets.In this process,we summarize the characteristics of point cloud data,and point out the quantitative gap between the current point cloud dataset and the two-dimensional image dataset.The quantitative gap limits the performance of object detection algorithms based on supervised learning.In order to make full use of point cloud data and enhance the feature extraction capabilities of deep learning networks,we refer to some self-supervised learning methods,and propose to create reconstruction scenes by rotating some specific objects in the original scenes.Afterwards,due to the simple structure of the single-stage object detection algorithms,we first design a single-stage object detection algorithm based on reconstruction scenes.The algorithm mainly consists of two tasks including the main task and the pretext task.The main task completes the prediction of the position and category in the original scenes,and the pretext task is used to predict the angle difference of the rotated objects in the reconstructed scenes.The pretext task and the main task share some features in the region proposal network.The pretext task uses the targets of the reconstruction scene to extract the difference features of the rotated objects from the shared features,and then enrich the information of this part of feature while completing the pretext task.And the pretext task enhances the feature extraction capability of the detection network and contributes to the improvement of the main task.Although the above single-stage method is simple and straightforward,it still exist a problem that the part of shared feature does not directly affect the detection results.Therefore,we further design a two-stage object detection algorithm based on reconstruction scenes.Different from the above single-stage method,the pretext task needs to match the proposals of the rotated objects first,and then extract the difference features of the rotated objects.And the difference feature and the two-stage fine-tuning feature are all derived from the feature extraction network,which helps to enhance the influence of the pretext task on the main task and improve the detection accuracy.We conduct experiments of the above two methods on the KITTI dataset,and the results also prove the effectiveness of our methods.We submit the best performance to the KITTI online test platform for inspection.And our result achieves the 3D average precision of 81.84%,which ranks the 14th in the moderate difficulty detection of vehicles on May.12th,2021.Finally,in order to verify the performance of our algorithm on the ship target in the sea scene,we collected and produced a point cloud dataset of the ship self-propelled model.After theoretical analysis,the object detection research on the dataset of ship self-propelled model is equivalent to the research on real ship data.Then,we use some existing methods and our twostage object detection algorithm based on reconstruction scenes to experiment on this dataset,and the detection accuracy of our method can achieve 81.33%average precision in the moderate difficulty,which is a 6%improvement over other methods.