The Research Of Vehicle Tracking Algorithms In Traffic Scene

Intelligent transportation solutions have become an effective means to alleviate traffic pressure. The vigorous development of artificial intelligence and Big Data has brought new vitality and development to the intelligent transportation field. Meanwhile,the intelligent algorithms based on traffic monitoring represent the future of traffic information acquirement attracting many attentions. As the basis of and the key to traffic information collection, vehicle detecting and tracking based on video provides basic data for acquiring various traffic parameters. Due to the complexity of the traffic scenarios, there are many challenges in vehicle detection and tracking, such as occlusions, illumination variation and the real-time requirement. In recent years, deep learning has made a breakthrough in computer vision, making the video analysis method being applied for traffic analysis and understanding become possible. With the extensive application of traffic monitoring equipment and the blowout growth of video data containing abundant traffic information, the rapid and accurate vehicle detection and tracking method have great significance to the traffic information acquisition and traffic management.This paper is based on various traffic surveillance scene in actual road. The state of art object detection and tracking algorithms are deeply researched. A thorough research is conducted to design an efficient and accurate vehicle tracking algorithm and a multi-module fusion vehicle tracking framework is proposed in this study. This study can be summarized as followed.Firstly, a vehicle detection method is proposed based on deep neural network. A robust vehicle detector is trained using deep learning framework Caffe according to the information of the vehicle and the surroundings.And the network structure of detector is constituted by a region proposal network and a detection network. Experiment shows that the algorithm has achieved excellent results in various weather and traffic scenarios.Secondly, an updating method of tracking model based on incremental learning is proposed to overcome the shortcomings of Kernelized Correlation Filters concerning model updating. A set of snapshots which containing early tracking models is established. The model updating scheme is thus designed through incremental updating,which considering the principal components of both the information of the early and present models. The robustness and adaptability of this technique is proved by the comparison with other updating methods.Finally, a vehicle tracking algorithm is proposed based on the vehicle detector and correlation filters tracking means. The detector is used to initialize the multi-scaled tracking method and correct tracking failure. The correlation filters method is adopted in tracking single vehicle in short term. The trajectory association method based on tracklet confidence is applied to realize the organic fusion of detection and tracking to obtain more complete vehicle trajectory in complex traffic environment. Experiments shows that the tracking framework achieves a balance between speed and accuracy and obtains excellent results on the field monitor videos.