Research On Vehicle Detection Algorithm Of Multi-Sensor Information Fusion Based On Deep Learning

With the rapid growth of car ownership and the complex development of driving road environment,traffic safety has become the focus in many countries.As one of the most important traffic subjects,the driving safety of automobiles plays a vital role in road traffic safety,and intelligent driving technology can effectively reduce the risk of human accidents.Therefore,the research on automobile intelligent driving technology has become an essential part of automotive field.Environmental perception technology,as the "eye" of the intelligent driving system,provides more accurate environmental information for the decision-making and control system,which is of great significance for improving the safety and efficiency of intelligent driving.Focusing on the fusion of visual information and lidar information,this paper researches on the vehicle target detection technology based on vision and lidar information,and improves the detection accuracy from the perspective of sensor information fusion.The main research contents are as follows:(1)Research on two-dimensional vehicle target detection algorithm based on visual information.Considering the accuracy and real-time requirements of the intelligent driving system,the algorithm is based on the YOLOv5 network model,the SwinTransformer feature encoding module is used to replace some of the CSP network structural blocks in the YOLOv5 network,which make progress in extracting shallow local features with a lower computational cost.Secondly,the Bi-FPN feature fusion network framework is used to replace the original PANet network structure,and the CBAM attention mechanism is used to enhance the performance of important feature fusion.Finally,the ST-YOLO algorithm is constructed,which improves the detection performance of the vehicle algorithm,especially the detection accuracy of long-distance small target vehicles and occluded vehicles,which lays an effective visual detection foundation for the later multi-sensor fusion.(2)Research on 3D vehicle target detection algorithm based on lidar information.Because the detection accuracy is not high enough when the Point Pillars algorithm converts point cloud data into pseudo-image data.The CBAM attention mechanism and the residual network structure block are introduced into the feature fusion network of Point Pillars to strengthen the fusion effect of the network on key features,thus build the CBAM-Point Pillars algorithm model.After experimental verification,the CBAMPoint Pillars vehicle detection algorithm model has higher detection accuracy than the original Point Pillars.(3)Research on vehicle target detection algorithm based on fusion of vision and lidar information.Aiming at the limitation of single sensor information,in order to make full use of vision and lidar detectors,a decision-level multi-sensor information fusion detection framework is proposed.The residual convolutional neural network structure is used to fuse the 2D and 3D detection results,outputting the fusion classification score,which improves the vehicle target detection accuracy.Quantitative experimental verification proves that the proposed decision-level fusion detection algorithm can accurately detect the forward vehicle,and has better detection accuracy than the vehicle detection algorithm based on a single sensor.