Research Of Preceding Multi-Vehicle Tracking Based On On-Board Monocular Vision

Multi-vehicle tracking plays an important role in traffic scenario analysis,such as Adaptive Cruise Control(ACC)and Forward Collision Warning(FCW)in Advanced Driver Assistance Systems(ADAS).As the basis and premise of the safety of the intelligent vehicle,the real-time motion perception of the vehicle in the front view can roviding reliable information for the vehicle’s more advanced behavior selection,which be realized by tracking.In this paper,a significant preceding multi-vehicle tracking method based on monocular vision is proposed,which using high-precision vehicle detection,multi-dimensional feature matching methods,and the deep learning tracking strategy to achieve real-time robust multi-vehicle tracking in complex traffic scenes.The main contributions of this thesis are threefold:(1)Based on the study of the single-stage detector,and an offline-trained vehicle detector is customized to generate robust and fine detections by an onboard monocular camera.To begin with,self-collectted dataset is collected in three typical traffic scenes.Image augmentation is adopted to improve generalization.The efficiency of the vehicle detector,named Car Hunter,is improved by residual network,multi-scale detection,and iterative training.The experimental results show that the proposed Car Hunter algorithm can effectively handle multi-vehicle detection under complex traffic scenes and lay a good foundation for the follow-up tracking system.(2)A vehicle association method based on multi-dimensional feature fusion is proposed in this part.The multi-resolution features of the appearance of the vehicle are extracted by the central-surround two-channel spatial pyramid pooling network,and the similarity of appearance is combined with the motion model to establish the similarity matrix.The Hungarian algorithm is introduced to find the best matching for the dictum to complete the association of consecutive frames.The experimental results show that the well-designed association strategy could improves the accuracy effectively and can still meet the real-time requirement.(3)In order to deal with the problem of frequent entry and disappearance of in-transit vehicles in the field of view,a method of managing the lifetime of each vehicle via the Markov model is proposed,which can handle the false positives in the tracking process.Inspired by the status characteristics during tracking period in real traffic scenes,the states of vehicle can be divided into four states: Probationary Period,Tracking,Vanishing,and Lost.Q-learning is adopted to learn the best state transition strategy,which can be used to improve tracking efficiency via manage the effective information by using appropriate switching of the status of each vehicle.Compared with other methods on KITTI and self-collected datasets with onboard NVIDIA Jetson TX2,the vehicle tracking method “Vehicle of Interest” achieves significant performance in terms of the “Mostly-tracked”,“Fragmentation”,and “ID switch” variables.Besides,the robust multi-vehicle tracking could meet real-time(15.3fps)requirement.