Research On Detection And Recognition Of Vision-based Human Activity

Vision-based human activity detection and recognition has been rapidly developed into a research hotspot in the domain of computer vision and artificial intelligence because of the increasingly demand. It has a wide application in the field of intelligent video surveillance, virtual reality, motion analysis, smart home, user interfaces and so on. Because human body is non-rigid object with too much degree of freedom, the human model sometimes can not precisely describe the real human body. The immaturity of the field of computer vision and complicated environment bring many challenges to research on detection and recognition of human activity.Based on the analysis of the prevalent algorithms of the detection and recognition of the human activity, this dissertation goes into a deep research on detection of motion human, feature extraction and activity recognition. The main contributions of this dissertation are as follows:1. After accessing and summarizing the relevant literature, we propose a moving object detection algorithm based on improved mixed of Gaussian models and symmetric difference algorithm. This algorithm based on mixed of Gaussian background subtraction to promote the learning rate of the background models, overcoming the disadvantages in the start of the building of the background models. Furthermore, with the integration of the symmetric difference algorithm, it can detect the moving object. Experimental results demonstrate this algorithm can detect the moving object real-time accurately.2. Importance and critical of the feature selection and extraction of the human activity recognition is introduced by bringing a shape description operator which called R transform. R transform is insensitive to the noise and cavity and it is robustness. R transform is invariant under translation and scaling but variant under rotation. We further promote the R transform's rotation invariant to make it better applied in the description of human activity feature. The principle component analysis (PCA) algorithm is used to reduce the dimensionality which is extracted by improved R transform.3. The hidden Markov model (HMM) which is effective to solve sequence signals, is used as classifier to recognize human activities. Experimental results show that the improved R transform has higher recognition rate than original R transform. Human activity recognition rate was increase by using one state hidden Markov model which can overcome the shortcoming of multi states hidden Markov model's complexity and inaccurate of the state transition probabilities.4. In order to check the contents mentioned above, a vision-based system prototype is designed which can detect and recognize the human activity. This system prototype has functions of video recording and reading, moving object detection, feature extraction and human activity recognition.5. At last, this dissertation is summed up and some expectations on further study are put forward.