Research On Object Detection Method Oriented To Coorerative Vehicle Infrastructure System

When the development of autonomous driving technology has arrived at L4 or L5,it is difficult to implement due to the impact of costs and other factors.It is imperative to develop intelligent infrastructure.The Intelligent Vehicle Infrastructure Cooperative Systems(IVICS)represents a new direction for the development of the Intelligence Transportation System in the future.The Road Side Unit(RSU)is used to sense the environment in the IVICS,and the traffic scenes they faced are larger and more complex.Aiming at the problems of poor stability、easy misdetection and missed detection of existing object detection methods,this paper is based on the national key research and development project "Holographic Traffic State Reconstruction and the Test and Verification of Vehicle Group Coorerative Control" to study the fusion strategy of lidar and camera,and a 3D object detection method with both stability and precision is proposed.The main research contents are as follows:(1)Based on the analysis of typical application scenarios of the IVICS,combined with RSU sensing technology and data fusion theory,a roadside heterogeneous sensor spatial fusion model is established,which realizes the data alignment between lidar and camera.Aiming at the problems of poor stability and low precision of existing fusion algorithms,a multi-level fusion framework based on DCFP(Depth Completion-based Frustum Point Nets)algorithm and AVOD(Aggregate View Object Detection network)algorithm is proposed.(2)Aiming at the problem that the classic Frustum Point Nets algorithm does not make full use of the RGB information in the image data and the point cloud is sparse,an improved algorithm DCFP based on depth completion is proposed.In the frustum proposal module the original sparse depth map is depth-completed under the guidance of the RGB image,then the dense depth map could be got.Then the dense depth map is combined with the 2D detection results of the image for ascending projection to obtain the frustum point cloud.Finnally the3 D bounding box of the object could be obtained by 3D instance segmentation and bounding box regression.(3)Aiming at the instability of the serial structure of the DCFP algorithm and the large detection error of small objects,the point cloud and image data are input into the AVOD algorithm for secondary detection to obtain another 3D bounding box.The fusion method of3 D bounding box is used to subject and decide the 3D bounding boxes detected by the AVOD and DCFP algorithms,and the final detection result is obtained.The proposed fusion method is experimentally verified on public data sets.The results show that the proposed method can improve the precision of object detection.Compared with the F-Point Nets algorithm,the average precision(AP)for "difficult" object is increased by1.72%.Compared with the AVOD algorithm,the average precision for "difficult" object is increased by 1.97%.Comprehensive experimental results show that the fusion method takes into account both precision and stability,and has a good application prospect.