Research On Moving Object Classification Based On Shape Features

Video-based motion analysis aims at detecting, tracking and identifying moving objects, and more generally, understanding objects behaviors through analysis and processing image sequences with moving objects. Objects classification is an important aspect of video-based motion analysis whose research content is to classify moving objects into semantically meaningful categories, associating the correct object class label with the region of interest. Object classification mainly used for activity understanding in scene. Therefore studying multi-class nonlinear object classification is significant for the development of automatic video comprehension.The moving object classification of normal outdoor scenes based on static odd-camera is studied in this paper. On the basis of summarizing and analyzing actuality research and algorithms both here and aboard, an object classification algorithm based on shape features and support vector machines is proposed. It classes the objects detected in video into four familiar sorts: person, crowd, vehicle, bicycles. The main work is as follows:1. Firstly, video images are pretreated through segmenting regions corresponding to objects then tracking them after self-adaptive background subtraction and noise wiped. On the basis of pretreatment, features of moving object are extracted. Some simple and efficient shape features are defined which are fit for the variety of object shape.2. Multi-class SVM classifier based on small sample is built. Labeled samples are used to train the classifier. The problem of multi-class non-linear object classification can be solved better. The trained classifier can be used to class unknown object samples.3. Some means which improve the performance of object classification are discussed. Interval frame classification is proposed. In this thought, features extract and classification are treated at interval frames instead of each frame, which debases time complexity of classification algorithm. And the methods of using temporal consistency and scene-dependent features are described.Simulation results show that this method can class person, crowd, vehicle and bicycles preferably. And it has some adaptability in both object varying multi-angle and multi-pose, object with a few shadow and imperfect object extracting. The whole...