Monocular Camera-based Vehicle 3D Detection And Bird’s-eye View Tracking Algorithm Research

In the past 6 years,vehicle 3D object detection methods based on monocular camera have been extensively developed in the field of autonomous driving perception in computer vision,and have achieved certain breakthroughs in detection accuracy and real-time performance.Based on the depth distribution-based monocular image 3D target detection model,this paper conducts in-depth research on related algorithms,and proposes a monocular image 3D target detection algorithm based on a cross-view attention mechanism and a BEV target based on multi-frame spatiotemporal embedding.tracking algorithm.In view of the fact that most of the current 3D detection algorithms that adopt a single perspective will inevitably damage the integrity of the target information in the 3D space,resulting in inaccurate prediction results,a monocular image 3D target detection algorithm based on a cross-view attention mechanism is proposed..The first is to predict the depth probability distribution of each pixel through the depth distribution sub-network,and project the rich contextual feature information into the three-dimensional space to generate threedimensional voxel features;the second is to project the three-dimensional voxel features to different directions to generate the front view captured by the camera The multi-view features of perspective view(Perspective View,PV)and bird’s eye view(Bird’s Eye View,BEV).The information is fused into the bird’s-eye view feature map and the classification and regression tasks are decoupled in the prediction network;finally,the effectiveness and performance superiority of the algorithm are proved by sufficient experiments.Aiming at the problem that if the target features such as target overlap and motion blur are not obvious in the automatic driving scene,the target may not be detected and the tracking target may be lost.Based on the model in Chapter 3,this paper proposes an additional multi-frame spatio-temporal The target tracking algorithm under the embedded BEV perspective,compared with the traditional target tracking algorithm under the camera perspective,on the one hand,this method converts the tracking perspective from PV to BEV,and the target tracking under BEV can alleviate the occlusion and overlap of targets under PV The interference caused by the situation to the tracking stability;on the other hand,this method introduces timing information,and through the fusion of timing features,it can alleviate the visual overlap and occlusion problems to a certain extent;accuracy and stability.Based on the research results of this paper,a traffic safety early warning system oriented to the road environment is implemented.The vehicle target detection system detects and tracks the road conditions in front of the camera captured by the camera.When the distance to the target vehicle is too close,the system uploads the early warning information in time.To the early warning information management system server.The early warning information management system can collect and manually check the uploaded early warning information.