Research On Vehicle Tracking Technology Based On Tegra Embedded Platform

Intelligent Transportation System(ITS)integrates computer technology,traffic engineering,communication technology and other technologies to provide convenience for human travel.Video vehicle detection and tracking is the core of intelligent traffic monitoring system.This thesis designed and implemented the vehicle tracking system based on Jetson TX2 embedded platform.The video is captured by SDI camera,and the video data is converted into YUV420 M format,and the vehicle is tracked according to the obtained dynamic video,then the tracked video data is sent to HEVC encoder for encoding,and the output code stream is encapsulated by RTP and sent to PC through UDP protocol,and the tracked video is received,decoded and displayed by Gstreamer on PC.When the vehicle in video is tracked,the vehicle detection and tracking algorithm is applied.Based on the traditional vehicle detection and tracking algorithm,the design is modified in the following three aspects:(1)Improved algorithm of multi-scale detection: In order to solve the shortcoming of the general framework of target detection in natural scenes,a multi-scale real-time Faster R-CNN target detector is proposed,and the multi-scale detection technology is introduced into the Faster R-CNN algorithm.Compared with the traditional Faster R-CNN detection method,its architecture is simpler and its model is smaller.The experiments of VOC 2007 dataset,Kitti dataset and real traffic dataset are carried out on Jetson TX2 using Tensor Flow framework.The results show that the accuracy is increased by 15.4% on average and the recall rate is increased by 13% on average.(2)Improved tracking algorithm based on predictive trajectory technology: Aiming at the problem that the traditional KCF algorithm(Kernel Correlation Filter)cannot deal with the scale change of the target effectively in the tracking process,this paper integrates the scale optimization mechanism into the tracking process,and combines the Kalman filter algorithm to predict the vehicle trajectory,complete the correction of the trajectory information,and realize the continuous tracking of the vehicle target.Experimental results show that the accuracy of the proposed algorithm is improved by 23.9% compared with the traditional KCF algorithm.(3)Aiming at the problem of speed,this paper uses Tensor RT to make rapid and efficient deployment reasoning for Tensorflow framework.The experimental results show that the speed of the algorithm is increased by 6.3 times after the optimization of the learning framework.