Chaotic Feature Vector Based Dynamic Texture Recognition

Dynamic texture recognition is one of the fundamental problems in computer vision and widely used in military and urban field. In this dissertation, in order to characterize dynamic texture accurately, chaotic feature vector is employed. Optimization algorithm is used to improve the dynamic texture recognition accuracy. Deep learning technique is used to learn the low level feature to a high level feature so as to enhance the recognition performance. Meanwhile, the dynamic texture recognition algorithm is applied to traffic video classification. The contribution of this paper is:1. Pixel intensity series is proposed for dynamic texture description as a basic feature. The pixel intensity series is treated as an integral that can be obtained more temporal information of the time series. The experiments indicate that the pixel intensity series possess the ability to describe the dynamic texture.2. Chaotic feature vector is proposed to descript the pixel intensity series based on the self-similarity exists in the pixel intensity series. The proposed chaotic feature vector is different from the previous chaotic feature vectors used in computer vision. In dynamic texture, the fractal property is important. The proposed chaotic feature vector which is a local descriptor can be used to characterize the dynamic texture. Combining with the bag of words model, chaotic feature vector based method can achieve better performance than the state of the art methods.3. The content based dynamic texture recognition framework can overcome the quantization problem that exists in the bag of words model. In this chapter, the pixel intensity series is represented by the chaotic feature vector. The segmentation algorithm is employed to obtain the foreground and background information. Then comparing the segmentation results between two dynamic textures with earth mover’s distance, the dynamic texture recognition is achieved. Furthermore, the proposed algorithm can be used in traffic video recognition.4. A novel recognition algorithm that considers the relationship between features is proposed. And alternating direction method of multipliers(ADMM) optimization algorithm is employed to solve the problem. The multi-task learning approach captures the features’ relationship: common information and individual information.5. A novel learning algorithm based on deep learning is proposed to learn the low level feature to a more compact and representative feature. The proposed algorithm can learn the middle level feature to a high level feature. The experimental results show the method achieves higher recognition rate.