Research On Video-based Vehicle Detection And Location In ITS

Intelligent transport system (ITS) is an important application of video surveillance technology, which will be the future tendency of this area. Though in the bottom layer, video-based vehicle detection and location technologies play key roles as ITS data resource. Their results accuracies decide the following processes effect, including vehicle tracking, feature extraction, image understanding and etc. So they are very important in ITS. At present background subtraction is used most, whose key step is background modeling. Though Gaussian Mixture Model can describe the traffic scene well, it can't meet the real-time requirement. The vehicle motion status change often causes a part of background invalid, so it needs a special approach to update the background model. There are not so many researches about vehicle location technology that the location accuracy is not good.According to the work flow of background subtraction, this paper mainly studies these topics: image pre-processing, background modeling and updating, vehicle extraction, image post-processing and vehicle location. By systematically analyzing former researches, this paper makes improvements about Gaussian Mixture Model (GMM) algorithm, background updating and region growing algorithm to satisfy the traffic specific requirements.The major works and contribution of this dissertation are as following:(1) In image pre-processing, Retinex algorithm is introduced to eliminate the traffic image noise causing by weather, sunshine and shaking trees. In terms of image contrast and definition, experiment results show local Retinex is best relative to histogram equalization and global Retinex.(2) In order to meet the real-time requirement, an improved Gaussian Mixture Model is proposed, which dynamically selects gauss model number for every pixel. Experiment results show that the improved approach can improve efficiency significantly. This paper proposed a local background updating method to overcome part of background sudden change problem, which uses image gray-relation to distinguish static objects and false objects. Experience results show that this approach can adjust background model to the changes.(3) This paper uses Otsu algorithm to threshold the difference image between current frame and background image. Often the binary image has isolated noise points or objects with holes, which are post-processed by morphological filtering algorithm. Experiment results show that vehicles are extracted correctly and without noise.(4) In the vehicle location step, watershed and region growing algorithms are invalid when the spatial domain connectivity of binary image is bad. Comprehensive consideration of accuracy and real-time, an improved growing algorithm is proposed, which uses a region of pixels to expand the range of influence and connects the far right pixel to the object. The vehicles are labeled by rectangles, whose cenroids are associated to track vehicles.