Algorithm Research On Video-based Vehicle Detection, Tracking And Recognition For Intelligent Transportation

The thesis mainly discusses the fundamental theories and key technologies of video-based detection, tracking and recognition of vehicle monitoring system for intelligent transport. Traffic detection and collecting information collected have become an important topic in intelligent transportation system, and moving vehicle detection, tracking, recognition background modeling and updating, and shadow removal are the essential basic modules. To the above problems, further study is done in this thesis and new algorithms are brought forward combing the original algorithm characteristics and application technology, experiments are implemented to demonstrate the validity. The main contents of the study include such aspects as follows.Firstly, object detection technologies are investigated. A real-time moving detection algorithm based on combing gray correlation frame differencing and background differencing is proposed.At first, Gray relevant frame differencing algorithm is proposed. Meanwhile, methods of background modeling are investigated, and an adaptive background updating method based on light change is proposed.Then shadows in the HSV space of the original color images are detected and removed. Finallyl, logical AND Operation with two difference algorithms are done, so the moving target is detected accurately. Experiments show that the proposed algorithm can eliminate noise, shadow, and can adapt to different illumination condition detection scenes, therefore it has higher accuracy with robustness and adaptability.Secondly, object-tracking technologies in complex background are introduced. A algorithm based on combing gray and contour feature information improved particle filter tracking, using democratic fusion strategy is proposed. In observational equations, a particle adaptive weights updating method based on gray and contour feature is put forward, using democratic integration strategy. Finally, the posterior probability density are calculated, and the block between vehicles to track the moving vehicles are handled. Experiments prove that this algorithm makes full use of the gray and contour features in different time under different scenarios for the different weights, to a certain extent, the block between vehicles are removed, so it has high accuracy and robustness for vehicles tracking.Finally, vehicles classification is investigated. The length, width, height, area, first, second moments from video manually as features of vehicles classification are extracted. One-to-one kinds of support vector machine(SVM) is used to classify vehicles, then the fuzzy support vector machine (FSVM) is used to classify the model of vehicles.Experimental results show that, because of considering the membership information of samples, the fuzzy support vector machine (FSVM) has better classification accuracy than the traditional support vector machine(SVM).