Research And Implementation Of Vehicle Target Tracking Algorithm For Intelligent Transportation System

Intelligent transportation system(ITS)aims to make full use of advanced information technology,sensor technology,artificial intelligence technology to realize real-time,efficient and accurate integrated transportation and management system.One of the key technologies is vehicle tracking.Tracking of vehicles using artificial intelligence and image analysis technology can not only save the cost of installing radar and other tracking equipment,but also provide data for the subsequent high-level tasks such as vehicle flow detection.However,due to high appearance similarity among vehicles and heavy occlusion caused by busy traffic flow,the accuracy of tracking has always been low.At the same time,most of the tracking algorithms combined with artificial intelligence rely on high-performance hardware equipment,and in the ITS is more composed of a large number of embedded devices,so the existing algorithms are not suitable for direct application in the power limited embedded devices.In order to improve the tracking accuracy and ensure a better effect in the actual complex monitoring scene,and explore the feasibility of the vehicle tracking algorithm on the embedded device,this thesis mainly conducted the following researches:1.Vehicle tracking is one of the important technologies of ITS,which commonly follows tracking-by-detection strategy.A major challenge in such a tracking system is the limited performance of the underlying detector which may produce noisy detections.Consequently,siamese network and backward prediction-based vehicle tracking approach is proposed.Siamese network based forward position prediction is designed to alleviate the interference of noisy detections,while backward prediction verification is performed to reduce the false positives arising with forward prediction.The final tracklets is obtained through weighted merging based on the detection confidence and forward prediction confidence.The proposed method is evaluated on a diverse set of benchmarks including UA-DETRAC and MOT17.The experiment results demonstrate that the proposed method outperforms the state-of-the-art on the UA-DETRAC vehicle tracking dataset,as well as maintains real-time processing at an average tracking speed of 20.1fps,which can be used for real-time applications.At the same time,the proposed method can maintain high processing speed and the performance is close to the state-of-the-art on MOT17 dataset.2.Based on NVIDIA Jetson TX1 embedded platform,the thesis designs and implements the vehicle tracking system.The system is mainly divided into two modules.The first part is the vehicle detection module based on the YOLO(You Only Look Once)detection algorithm,and gives the specific process.The experimental results show that the YOLO algorithm can effectively identify the vehicles in the traffic monitoring video.The second part is the vehicle target tracking module based on the vehicle tracking algorithm proposed in this thesis,and it is compared with a tracking algorithm based on global optimization on the Jetson TX1 embedded platform.The experiment shows that the algorithm proposed in this thesis can also maintain high tracking accuracy and tracking speed on Jetson TX1,which proves the feasibility of the algorithm running on the embedded device.