Researches On Monocular 3D Object Detection Based On Pseudo-LiDAR Point Clouds

Huge challenges have been brought to urban transportation on account of the global urbanization.Recently,automatic pilot has oriented the development of Smart Cities as the rapid rise of artificial intelligence.3D object detection algorithms perceive the environment with the on-board sensors,such as camera,Lidar or ultrasound.and obtain parameters of the surrounding vehicles to ensure the safety of automatic cars,which makes 3D object detection become the focused areas of automatic pilot.Monocular 3D object detection attract attention by its low cost.However,it is difficult to obtain accurate depth information from monocular images due to the limitation of the camera,which leads to accuracy loss when detecting the position or size of the objects.This thesis uses pseudo-Lidar framework for monocular 3D object detection.Pseudo-Lidar point cloud is generated form DORN(Deep Ordinal Regression Network)depth map and detected by PV-RCNN(Point and voxel RCNN).A high-performance monocular 3D object detection algorithm,PLIDP(Pseudo-Li DAR based on DORN and PV-RCNN)is proposed according to this pseudo-Lidar framework.To improve performance of PLIDP,the thesis proposes solutions for the lack of depth information in monocular images from the perspectives of object salience and the point feature extraction based on neural network.The research of this thesis as follows:(1)As the perspective of object salience,a layered structure-based confidence generation model is proposed to improve the quality of pseudo-Lidar.The model generates global and local confidence respectively.Global confidence is taken as the weighted value of local confidence.This model assumes that local confidence obeys two-dimensional Gaussian distribution in the 2D bounding box,while global confidence obeys the linear distribution that decays with the increase of the point cloud depth in the scene.The experimental results turn out that the salience of objects improves in the pseudo-Lidar point clouds resampled according to confidence,which also leads to the improvement of PLIDP.(2)As the point clouds feature extraction perspective,the Dynamic Graph CNN(DGCNN)is introduced,and a feature aggregation DGCNN model is proposed to improve the feature extraction ability of detection network to pseudo-Lidar point clouds.This model use DGCNN to extract geometric structure features of distorted object,and use multi-feature aggregation network to aggregate geometric features and color features.Experimental results show that this model can effectively improve the performance of PLIDP.