Research On Target Detection And Tracking Of Vehicle Platform Based On Improved CNN

Nowadays,cars bring convenience to people’s travel,but also bring traffic safety hazards.The driver’s inattention and fatigue driving often lead to traffic accidents.Therefore,the vehicle safety assisted driving system is of great research value.In these studies,vehicle target detection and tracking in the vehicle platform environment is one of the important research contents.First,the traditional detection methods are too dependent on manual adjustment,and the target detection effect is poor.Secondly,the traditional tracking method is in the target itself.When the change is severe or occluded,it is easy to drift.Finally,although deep learning has made breakthroughs in the video field,researchers can always design a powerful detection and tracking model with more parameters to digest the data,but a large number of weights will take up a lot.The storage bandwidth makes it difficult to deploy to the in-vehicle platform.Therefore,how to detect and track the vehicle in real time under the premise of ensuring high precision in the complex urban road driving scene and deploy the model to the embedded platform becomes an urgent problem to be solved.In this context,aiming at the shortcomings of existing vehicle detection and tracking algorithms,this paper proposes a real-time vehicle detection algorithm(SAVD)based on scale-aware CNN.Firstly,the existing Ro I pool has the original structure of destroying small objects.The scale-aware Ro I Pooling Layer(SARo I)generates accurate feature maps for small-sized vehicles.At the same time,this paper proposes a multi-branch decision network for vehicle detection.Each branch is designed to minimize the inter-class distance of features,so it can capture the distinguishing features with various scale targets more effectively than existing networks.Experiments show that the SAVD algorithm is robust to vehicles of different space sizes and better for small-sized vehicles.Aiming at the problem that the SAVD network model is large in size and consumes a large amount of GPU resources,it is difficult to deploy in embedded systems with limited hardware resources.This paper proposes an adaptive joint pruning-quantization(AJP-Q)based network compression algorithm.The SAVD network is compressed and accelerated.The detection performance of the model is evaluated on the TX2 embedded platform.The experiment shows that the compressed SAVD＿VGG network model can be deployed well under the embedded platform,and the running speed is greatly improved under the acceptable detection accuracy.Finally,this paper designs the geometry,target shape and attitude cost based on multi-target tracking in road driving scenes.With only the image of monocular camera,it can design the pairwise association of target trajectories according to several 3D cues such as target posture,shape and motion.cost.These costs are easy to implement and can be calculated in real time and complement each other.Finally,the online multi-target tracking is realized by the association framework of bipartite graph matching.Experiments show that the associated cost of this paper can improve the correlation between targets and help improve tracking accuracy.