Research On Feature Expression And Classification Algorithm Of Moving Object In Mixed Traffic

Transportation information collection system is an essential part of ITS (Intelligent Transportation System), it is the foundation of modern traffic management and control. At present, the traditional loop detector is widely used for traffic information collection, but disadvantages of loops caused by installation and maintenance restrict the detection precision and the detection range for collecting information accurately in real-time. With the rapid development of computer vision, image processing, pattern recognition and AI (Artificial Intelligence), video detection applied in the traffic area becomes a research focus. Video detection can obtain the state of moving objects and parameters of traffic flow for management and control of urban traffic. Compared with traditional information collection method, video detection has advantages in installation charge, detection range, function extending, useful life and so on.There has been lots of researches on video detection in the past, but most of them only take the vehicle as the study object. Researches existing can't be applied in mixed traffic flow, which is the mainly property of China traffic. It's necessary to study on video detection and video detection system for mixed traffic flow. According to mixed traffic property of our country, past research results are summarized up, the research meaning is defined, and crucial techniques of video detection are studied, including image pretreatment, moving object feature expression, moving object recognition and classification, which is combined with lots of intelligence algorithms in this paper.Image pretreatment is the basic part of video detection. The purpose is to extract moving objects from image sequence completely and accurately. Background Model can realize foreground objects detection under the influence of the variation caused by the environment, including initialization background, background expression and background update. Initialization background is the precondition to obtain the Background Model. The paper presents a new background initialization algorithm based on clustering classifier. In the algorithm, all stable sub-intervals in the training sequence are located for each pixel as possible background firstly, and a classify data set is constructed by the median values of each stable sub-interval, then a background sub-set is obtained from the classify data set by unsupervised clustering. Accordingly, the initializationbackground is obtained by the background sub-set. Combined with Mixture Gaussian Model based on the object level, it also presents an improved background update algorithm. In the experiment, different traffic condition videos are tested, and results show that proposed methods are robust and self-adaptive.After eliminating background by Background Model, the extracted foreground consists of moving objects and moving shadows sometimes. The moving shadow has a great impact on feature expression as well as object classification. In order to increase the detection validity and the recognition correctness, the paper presents a effective shadow remove method: twice-segmentation. At first, segmentation is used on the color image after foreground object extraction, the moving shadow of each foreground object is eliminated by its color information, and then the color image is changed to the binary image, the second segmentation is applied on the binary image to realize twice-segmentation. Shadow detection algorithm based on RGB vary degree of each object is used between twice segmentations to overcome the influence of moving shadow effectively.Feature expression isn't only related with the classification machine design and capability, but also has a full impact on object recognition and classification system. According to mixed traffic property of our country, two simple effective feature expression methods are presented in the paper based on discussing feature expression of vehicle and non-vehicle, one is feature expression based on the horizontal projection, the other is based on centro-distance vector. The results show that the feature expression based on the horizontal projection is simple and valid, the feature of one class is obviously different from others; the centro-distance vector feature is fixed on rotation, translation and scale, it can be applied for mixed traffic flow due to its stability and adaptability.In this paper, analysis is conducted on classification algorithms from real application, SVM (Support Vector Machine) adapted to a few samples is deeply studied to design classification machine, moving objects are divided into three classes: the vehicle, the bike and the person. The multi-class nonlinear problem is well solved by SVM. Vote mechanism is used in the process of object motion tracking to enhance classification reliability. It is well done for classification between vehicle and non-vehicle.In the end, findings and achievements are summarized up, and issues for further research are brought forward.The paper research provides the theory and algorithm for the intelligenttransportation information collection system, it also supports the management and control of mixture traffic on technique. The research findings of the dissertation have certain theory significance and practical value to impel ITS modernization advancement of our country.