Research And Implementation Of Behavior Recognition Method Based On Visual Perception

Behavior recognition is an interesting and challenging research subject in the fields of computer vision and pattern recognition, and it is the advanced processing link of visual motion analysis and understanding, which belongs to the harder visual task. Human action recognition has profound meaning of theoretical research and wide application prospect. It can be widely applied in the fields of security surveillance, virtual reality, motion skill training and medical diagnosis. Human action recognition based on computer vision recognizes human actions intelligently through understanding and analyzing the feature descriptor of human actions, which is on the basis of successful implementation of motion tracking and feature extraction. Currently, almost all the research just stay in the phase of simple actions, moreover, human body posture and movement are diverse, and the camera position and the real-life environment is so complex and so on, which make the action recognition still a hot subject in the research fields of computer vision.Given the above problems, this thesis learns the features using the image processing mechanism of human visual perception, after summarizing the state-of–the–art methods in recent years. The main research contents of the thesis follow:1. Elaborating the neuro-physiological principle of visual perception in human visual system in detail: how to extract the feature, and the characteristics of visual cortical neurons; and researching deeply the features learning algorithm based on the visual perception model.2. The thesis chooses the ISA model by the comparison of feature learning algorithms based on visual perception model. In the thesis, it uses a new learning algorithm of ISA model—relative gradient, which greatly improve the efficiency of model. In addition, we propose a new feature learning algorithm-the stacked convolution ISA algorithm.3. Elaborating the action recognition system this thesis uses in detail. The overall system framework includes the sampling, the data preprocessing, feature extraction, feature vectorization and the final classification. And exploring and studying the methods and their implementation in each processing module of the framework.4. System simulation is carried out on this action recognition system using MATLAB. We evaluate the algorithm in this thesis on the well-known benchmark action recognition dataset video Hollywood2, and compared with some advanced methods in this subject. The result shows that the action recognition method based on visual perception can learn effective salient features, achieve higher recognition accuracy rate.