Research And Realization Of Human Behavior Recognition Based On Fast-moving Scale Invariant Feature Learning

Human behavior recognition is a challenging computer vision task, it has been widely used in intelligent video surveillance, human-computer interaction field and motion analysis. Traditional human behavior recognition methods are based on ordinary video. With the widespread use of Kinect, Prime Sense, Leap Motion and other somatosensory equipment, Kinect-based human behavior recognition technology have been into the spotlight. In this paper, on the basis of using the database collected by Kinect, we focus on the key technology research on the human behavior recognition from the theoretical point of view. Our work includes the following aspectsIn this paper, we lead to the advantages of the Kinect based on the principle of the depth image achieved by Kinect. We set the widely used target detection method used in the ordinary video to the depth video, and conducted a confirmatory analysis to it. To the database of with hand gestures, we proposed a hand region segmentation method based on the rate of change of edge. The above pretreatment process is the foundation of feature extraction and description.For the contribution of the human behavior recognition with Kinect, a large number of studies have focused on the use of bone models, but the use of depth information in human behavior recognition has relatively little research. In this paper, our fast-moving scale invariant feature extraction and description method has combine the existing optical flow method, SURF (Speeded-Up Robust Features), and depth information, make an improvement of the two-dimensional information characterization. For fast moving of target, using optical flow updating method based on database to prevent the mismatching problem caused due to excessive movement. In this paper, our method can improve the feature extraction and description accuracy.Human behavior classification is a critical stage after feature extraction, it plays a key role in the recognition results. The widely used SVM (Support Vector Machines) has the disadvantage of poor generalization ability. In this article, we combined the SVM algorithm and HMM model, proposed a multi-level content-based classification method. We formatted a top-down decision tree based multi-level SVM classifier through judgment rule. Then, we use the HMM model for further classification.Based on the above work, we design and implement a comprehensive human behavior recognition platform. The platform is for online testing and analyzing the method proposed in this paper.