Vision-based Motion Human Feature Description And Behavior Recognition Research

Vision-based human behavior recognition is a hot research orientation in the fields of computer vision and machine learning. It is widely used in many applications, such as intelligent surveillance and human-computer interactionIn this paper, a deep research has done on the steps of motion target detection, feature extraction and description, and human behavior recognition. The main work and achievements are as follows:1. In the step of motion target detection, based on Visual Background extractor(ViBe), Robust Adaptive Visual Background extractor(RAViBe) is proposed in the paper. RAViBe uses a robust background modeling method based on image contrast, uses an adaptive background update strategy and adaptive target classification algorithm to detect motion target. The experimental results show that, compared with the original ViBe, the RAViBe can get better detection result in motion target detection, can quickly and efficiently eliminate ghost area, and can guarantee the algorithm robustness in the case of illumination change.2. In the step of feature extraction and description, the paper proposes a Block Based Histogram Fusion Feature Descriptor(BBHFFD), incorporates the silhouette feature with optical flow feature into the fusion feature to describe human behavior, and uses a block based histogram descriptor to descript the fusion feature. On this basis, the paper uses the method of Discriminative Common Vectors(DCV) to reduct the fusion feature dimension. Because the algorithm in computing performance is not high enough in the process of classification problems, the paper proposes an improved Fast Discriminative Common Vectors(FDCV) method, using scalar instead of vector to do the classification. Experimental results show that, compared with the single feature, the fusion feature is able to get higher recognition rate in the behavior recognition. Meanwhile, compared with DCV, FDCV can get about two times higher computing speed during the target classification with the guarantee of recognition rate at the same time.3. In the state space based human behavior recognition, the paper introduces the SVM-HMM hybrid model in human behavior recognition applications. Because of using Support Vector Machine(SVM) which is one of the discriminant models to replace the GMM module which belongs to the generative model, the SVM-HMM hybrid model obtain better classification performance while the training samples is insufficient. The experimental results show that, compared with the GMM-HMM, SVM-HMM hybrid model has better recognition performance in both short and long video sequences.4. In the rule based human behavior recognition, the paper proposes an evolved rule which is formed by a group of constraints for human behaviors.. The paper applies the genetic algorithm in the rules update, so that the rules can be constantly evolved. The experimental results show that, compared with the initial rules, the evolution rules can get higher recognition rate in human behavior recognition.5. Combined with the above research work, according to the application environment of hydropower stations abnormal behavior monitor, this paper uses OpenCV to implement an abnormal human behavior monitoring system in windows platform. After training, the system can recognize the abnormal behavior in fixed scene and display the alarm information.