Research On Human Behavior Recognition Based On Depth Sequences

Since the 1980's, Human behavior recognition in the field of computer vision had attracted wide attention of researchers. Among the current behavior recognition methods, Most of the problems are solved by using the traditional camera. Over the past decade had witnessed the rapid development of 3D data acquisition technology, especially after the emergence of Microsoft's Kinect, depth image began to be widely used. Using depth sequence for human behavior recognition can overcome the influence of the factors such as illumination change, occlusion, and environmental changes and so on. Therefore, this paper mainly study human behavior recognition based on depth sequence, and focus on the extraction of depth behavior characteristics.Aiming at the problem of low accuracy of traditional behavior recognition, method of combining depth image sequence with RGB image sequence, the HOGD features extracted from the depth image and the LBP features extracted on the visible image are fused to get the HOGD-LBP feature. This fusion feature overcomes the problem which HOG features only extract the geometric edge information of the human body and ignores the flat surface. Experiment results show that the recognition accuracy of fusion feature is better than the single feature.Aiming at the problem of low accuracy of human behavior recognition based on depth video sequences, extracting a new space-time local normal(STLN) feature to describe the behavior from the depth video sequence. STLN feature descriptors which use 4D spatial normal histogram can capture human motion and geometric information. We use polychorons to quantify the 4D space. Experiments show that the STLN descriptor can achieve better recognition results.Aiming at low recognition accuracy problem which bag of feature(BoF) coding can not describe feature space layout information and unable to locate the behavior object or capture patterns of behavior, a new locality-constrained linear coding(LLC) method is adopted to replace the bag of feature(BoF) coding. This method use the locality-constrained linear coding(LLC) standard to train codebook based on the bag of feature coding. In order to better describe the 4 dimensional feature space information, we use local constraints instead of sparse constraints, which can obtain locally smooth sparse and solve the optimization process of a large number of calculations.According to the depth video sequences, a complete behavior recognition system is established. Firstly, the local interest points are detected on the depth video sequence. Secondly, the interest points are described by using the 4D local normal vectors, and then the STLN are obtained. Then,in order to get the behavior representation a locality-constrained linear coding(LLC) is used to encode the features. Finally, the behavior representation is identified by using SVM.This paper does some research on the feature description of depth behavior recognition. Test on the DHA, MSRAction3 D, and MSRDailyActivity3 D depth databases. In this paper, the proposed algorithm can achieve better recognition results.