Vehicle Detection And Tracking Based On Night Traffic Video

Intelligent transport system (ITS) is currently being studied and worldwide focused on in the area of world traffic transport. One of the important contents in ITS is the vehicle detection and tracking in the traffic scene. It provides a key technology for the traffic detection and information collection. Meanwhile, in the field of computer vision and machine learning, moving object detection and multi-object tracking technology has been a hot topic. The core technology in the field of intelligent transportation field has a wide range of applications. Our research was conducted, on the basis of correlative works inside and outside China, to detect the vehicles in the traffic video at night and put forward some rule-based categorization strategies, in order to further determine the authenticity of the detected candidate lights. And then, on this basis for tracking and matching lights, we test the robustness of the night-time vehicle tracking. At last, we test and verify the accuracy of the method through the automatic evaluation technology.The main contribution of this thesis are as follows:1. Rule-based two-stage method of detection lights. Previous work we carry out the lights detected by the positive and negative lights selecting sample training Adaboost classifier. Considering the method cost too much time, their robustness are not ideal enough and so on. We Proposed a two-stage lights detection scheme based on rules. The first step is by improving the adaptive Otsu segmentation algorithm after multi-scale threshold segmentation lights to give the candidate region; The second step is based on the factor of candidate area to formulate decision rules, judge the detected candidate lights again and to get the final test results, In our collection of night traffic video data set, the two stage lamp detection method based on rule achieves 84% accuracy, and develop statistical sample size is much smaller than the rules require labeling classifier training required sample size.2. Vehicle tracking based on vehicle lamp tracking and matching and predicting the forward and backward trajectory of vehicle. This paper aims to use multi-object tracking technology to track lights which appearing in continuous video frame. And then the lights will be matched. Furthermore, a pair of headlights as a representative of the characteristics of the vehicle, completing vehicle tracking. In the process of tracking, in order to address the impact of the front and rear of the vehicle to block the acquisition of accurate information, proposing a two-way predicted the track fitting technique based on the speed estimation. It can solve the problem of occlusion vehicle during the short time. At last, based on evaluation accuracy rate, miss rate, false positive rate and conversion rate, the proposed method is verified the robustness.