Research On The Algorithms Of Moving Object Detection And Tracking In Video Sequences

With the development of science and technology, the application of computer vision is expanding and it has a great impact on human production and life. Target detection and tracking is an important research topic of computer vision, which relates to many areas such as image processing, artificial intelligence, pattern recognition and etc. Target detection and tracking has a wide range of applications on the areas such as unmanned vehicles, national defense and security, intelligent video surveillance, human-computer interaction based on video and etc. Therefore, research on this subject has great significance. This thesis mainly studies moving target detection and video tracking techniques. The main works include as following:1. This thesis describes corresponding basis of image processing. Firstly, introducing three kinds of color space and introducing conversion among them. Secondly, introducing four kinds of graying methods for color images:single color value method, average value method, weighted mean method and maximum method, then showing the process of graying an image with weighted mean method. Then, introducing two common methods of image filtering:mean filtering and median filtering and pointing out their advantages and disadvantages. Finally, we introduce the binary morphology:dilation, erosion, opening and closing.2. Research detecting method on moving target accurately from video. This is the key to success of subsequent processing such as target motion estimation, target motion estimation, target recognition, behavior understanding and etc. This thesis describes and analyzes some common methods of moving target detection including optical flow method, the inter-frame difference method and background subtraction method. This paper focuses on background subtraction and introduces several common background subtraction methods including statistical average method, median filtering method, single Gaussian model and Gaussian mixture model. Then, introducing the principle and algorithm implementation of these methods and pointing out their own advantages and disadvantages by some experiments.3. On the aspect of moving target tracking, we introduce some common methods including tracking method based on feature, tracking method based on the model and tracking method based on the active contour. Then, analyzing every method and pointing out their advantages and disadvantages. This paper focuses on the mean-shift algorithm, which is excellent and has characteristics of strong robustness, fast real-time processing and easy to implement. When the moving target has partial occlusion or larger rotation, the tracking result with traditional mean-shift algorithm will become inaccurate, even loss the target because of a single feature space of the traditional algorithm. To overcome this shortcoming, this thesis presents a tracking algorithm based on color and texture model. In the framework of traditional mean-shift algorithm, from multiple perspectives, matching features by fusing color model and texture model obtained from Contourlet transform are established.4. The proposed algorithm has been realized on an embedded platform. Experimental results show that the joint target model of color and texture is more effective and robust than target model of single feature space. On the condition that targets obviously rotate, even they is invisible, making use of the texture information, the algorithm can track the targets accurately.