Research On Heterogeneous Sensing Data Fusion Technology Based On LIDAR And Monocular

With the rapid development and promotion of digital city,the application demand of intelligent sensing system in automatic driving,natural environment monitoring,security and other fields are growing.Due to the limitations of single-mode sensor in capturing scene target information,it is necessary to use the complementary characteristics of sensor data of multi-source heterogeneous sensing system to reduce the uncertainty of single-mode sensing system,so as to realize the full coverage detection function of the surrounding environment.Therefore,in this paper,by combining the characteristics of LIDAR point cloud and monocular image,the data fusion method is studied based on deep learning algorithm,then a highly reliable and robust 3D target detection method is realized.The main research contents are as follows:(1)In view of the sparsity and spatial discrete distribution of LIDAR point cloud,a graph convolution feature extraction module is designed by combining voxel partition and graph representation,and a 3D LIDAR point cloud object detection algorithm in view of voxel based graph convolution network is proposed.By eliminating the computational redundancy of traditional 3D convolutional neural network,not only the object detection ability of the network is improved,but also the analysis ability of point cloud topology information is improved.(2)Aiming at the problem that the detection accuracy and the recall of single-mode object detection network algorithm decrease under different environment constraints such as climate,visual angle and light intensity,a 3D object detection method based on the fusion of the LIDAR point cloud and monocular image is designed based on the position encoding and channel-wise attention mechanism.The complementary characteristics of LIDAR point cloud and monocular image are fully utilized to improve the ability of sensing system to capture the spatial topology information and texture information of the target scene.(3)In order to verify the performance of the proposed object detection method,the 3D object detection and bird eye’s view object detection tasks of cars,pedestrians and cyclists are evaluated in the KITTI open data set.Experimental results show that,compared with the benchmark network Voxel Net,the single-mode LIDAR point cloud object detection method proposed in this paper effectively improves the object detection performance,especially in the 3D car detection benchmark,the average accuracy is up to 13.75%.Through multi-source data decision fusion network combined with image object detection information,the object detection accuracy is improved by 3.34% on average on the basis of LIDAR point cloud object detection method,and the perception ability and detection accuracy of single-mode object detection method are improved.At the same time,a visual interface is designed to evaluate the proposed target detection network.