Research And Application Of Obstacle Perception Technology In 3D Scene Based On Deep Learning

The development of science and technology as well as the breakthrough of key technologies will often solve important scientific problems in related fields,thus driving the vigorous development of the social economy.Thanks to the rapid innovation of deep learning technology in recent years,many problems that traditional technology cannot solve are constantly being overcomed.Innovations in hardware technology and the emergence of large-scale data have helped deep learning advance in the field of computer vision.In the era of artificial intelligence,Self-driving cars are the means of transportation for humans and their related technologies have extremely important research significance.This topic mainly proposes solutions for 3D scene understanding in complex road scenes in self-driving.Scene understanding technology mainly includes two aspects: detection and tracking.For the detection task,this thesis conducts an in-depth study on the 3D object detection algorithm based on deep learning and designs a comprehensive feature with local information in the point cloud data representation of LiDAR sampling.Then a deep learning model(SCANet)with autocorrelation attention mechanism is proposed on the 3D detection.For the tracking task,this paper uses the Kalman filter and the Hungarian matching algorithm to estimate the motion state of the obstacles and correlate the objects motion trajectory respectively.The main research contents and innovations of this paper are as follows:(1)In the existing 3D object detection technology,the information of point cloud data is not mined sufficiently.This paper proposes a new feature design method,which first transform point clouds to 3D voxels and then introduce the local orientation information to learn comprehensive point clouds representation by encoding the horizontal angle of each voxel.(2)In order to solve the problem of detecting small and occluded objects,this paper considers the point cloud data collected by LiDAR and the image information collected by the Camera.We design a new self-correlation attention module,which aims to fully exploit the space and appearance information of point clouds and images respectively as well as it could find complementary information to eliminate the modal gap.Finally,the results of the KITTI benchmark experiment demonstrate that the proposed method achieves state-of-the-art results for 3D object detection(3)To test and verify the proposed deep learning framework,it is applied to the detection of obstacles task in autonomous driving after the compression of the model.Then the Kalman filter algorithm and the Hungarian matching algorithm are used to estimate the motion state of these detected obstacles and correlate the motion trajectories respectively.We simulate multi-objects tracking on the ROS platform and the final rendering results demonstrate the effectiveness and robustness of our proposed model.