Research On Multi-scale Target Detection Method Based On Multi-sensor Information Fusion

Environmental perception is the key technology of autonomous driving.Intelligent vehicles use multiple sensors to detect and understand the surrounding environment,including various static and dynamic obstacles.In this thesis,aiming at the problem that the target in the near-field region of intelligent vehicle in the current environmental perception task is too close and the scale becomes larger,resulting in the outline beyond the perception perspective,changing from regular continuous shape to discontinuous natural shape,which increases the difficulty of target detection,a near-field region target detection method based on multi-sensor information fusion is proposed.The specific contents are as follows: Environmental perception is the key technology of autonomous driving.Intelligent vehicles use(1)Aiming at the difficulty of sensor data fusion caused by different sensor perspectives and different data representation forms,a multi-sensor feature information fusion method based on deep learning was proposed.The semantic category information of the image and the position and velocity information of the point cloud are fused to assist each other to improve the accuracy of near-field region target detection.Firstly,the central point image detection network was used to provide the initial region and semantic information of the target,and the 3D region of interest truncated cone is constructed.Then,F-Point Pillars method was proposed to fuse the feature information of lidar point cloud and radar point cloud to provide the target position and velocity information.Finally,according to the nearest neighbor principle,the image feature information and point cloud feature information in the truncated cone region are fused to finely regress the target state information.The comparative experimental results show that in Nuscenes dataset,the fused F-Center Fusion algorithm improves the NDS by 9.4 % and m AP by 3.5 % compared with the single camera detection method.(2)Aiming at the problem that the target contour of the near-field region in the environmental perception task is beyond the perception perspective,and each sensor can only obtain local target information,a multi-attention module is constructed to optimize the feature extraction module of the deep learning detector.The multi-attention module is added to the feature extraction network layer of the image and point cloud detection.By learning the different weight ratios of the feature map,the class dependence features can be quickly located,the feature expression ability can be enhanced and the invalid feature information can be eliminated.Secondly,the dynamic connection mechanism was used to fuse the feature information to solve the difference problem of different modal information between the color texture of the image and the geometric space characteristics of the point cloud.The feature information of different shapes and abstract levels is efficiently and learned,so that the detector can capture more discriminative feature information,and then improve the accuracy of target detection in the near field area.The experimental results show that in the Nuscenes dataset,the ADF-Center Fusion algorithm improves the NDS by 2.5 % and m AP by 2.9 % compared with the F-Center Fusion algorithm.(3)Finally,the effectiveness of the proposed method is verified based on the intelligent driving platform.The rotation matrix and translation matrix between the lidar coordinate system,the camera coordinate system and the radar coordinate system were obtained through the joint calibration of multiple sensors,and the spatio-temporal registration of sensors was realized.Then,a large number of real vehicle experimental data are collected and processed,and the scene data set suitable for nearfield region is produced to verify the effectiveness of the fusion algorithm in this thesis.The experimental results show that the fusion method can meet the requirements of near-field target detection accuracy.