Vision-based Human Movement Analysis

Human movement analysis is the most active research topic in computer vision. Its application in motion analysis, intelligent video surveillance, virtual reality and human-computer interaction is very promising. At the core of human movement analysis, it is critical to extract useful information from the video to detect, track and recognize human action. However, there are still some problems in practice, such as the performance of foreground detection in complicated environment, feature selection and extraction of reducing dimension and considering the sample information, Effectiveness of real-time classification of high dimensional nonlinear sample. Therefore, research on vision-based human movement analysis has important theoretical significance and practical value.We gain insights into some of the key technologies and problems of human movement analysis. The main contributions of this thesis are as follows:1. We summarize the representative algorithms on human movement detection. As an extension, aim to improve the performance of foreground detection in complicated environment, we propose a human movement detection algorithm based on background subtraction using codebook to update background model in real-time. Experimental results show that the proposed algorithm is effective in detecting movement in complicated environment; also it is robust against background disturbance, environmental fluctuations and other negative factors.2. We analyze the Principal Component Analysis (PCA) algorithm which is widely used in feature selection and extraction. PCA algorithm is essentially a linear scheme which is a limitation in application. Thus, we transform the samples into kernel space by nonlinear function and process them by PCA algorithm to preserve the nonlinear contents, in the meanwhile, we use a weighted method to enhance the contribution of the main feature in recognition. Experimental results show that the feature extracted by weighted kernel PCA could represent the sample information while reducing the same dimension compared to the traditional algorithm.3. Vision-based human action recognition is a typical problem of high-dimension nonlinear samples classification, the number of training samples directly affect the accuracy and timeliness of classification. We choose two types of classic classifier: K-nearest neighbor classifier and Support vector machine (SVM). Experimental results show that SVM had strong generalization ability for action identifying, which is a kind of classification problem for numberless and nonlinear samples, and able to achieve better recognition performance.4. Combined of movement detection, feature selection and extraction, action recognition, we design and develop a system for vision-based human action recognition, achieve various functions such as video recording and loading, human movement detection, feature selection and extraction, classifier training and recognition and demonstrate that it is a reliable experimental research platform.