Research On Vision-Based Human Behavior Recognition And System Implementation

Vision-based human behavior recognition is popular in the fields of computer vision and machine learning. It’s widely used in intelligent surveillance and human-computer interaction. Because of the complexity of scenes and the individual difference between same behaviors, researching on vision-based human behavior recognition is still a challenge.In this paper, we had a deep research on human motion detection, feature extraction and behavior recognition. The main work and results are as follows:1. ViBe is chosen as a rapid and efficient method for background modeling and foreground detection. But the distance threshold in ViBe is a fixed value, which is not adaptive enough for a variety of scenes. In order to solve this problem, OTSU method is used to produce an adaptive threshold, making the improved ViBe much more adaptable.2. With only appearance feature or motion feature, human behavior recognition cannot achieve high accuracy rate. On the other hand, state-of-the-art mixing methods are too complex. In order to solve this problem, this paper put forward a fast mixing method using silhouette and optical flow. The experiment shows that the mixed feature improved the recognition rate and the mixing method is more simple and fast.3. The widely used GMM-HMM is not good at recognizing short image sequences. So this paper introduced the SVM-HMM into behavior recognition field, which has proven to be successful in speech recognition. The GMM module in GMM-HMM is replaced by SVM, because SVM has better classification performance than GMM. So combining with HMM, the SVM-HMM has better performance for both short and long image sequences than GMM-HMM.4. According to the research above, this paper designed and implemented an intelligent surveillance system for campus. The system can recognize abnormal behavior such as fall and stoop in a fixed and side view.