Research On An Action Recognition Algorithm Based On Skeletal Joints

Human action recognition has very important application in human-computer interaction and video surveillance, and human action recognition is a very important and challenging issue. Most traditional action recognition is focused on video image sequences, but video image sequences are subject to interference of light, shadow and environment et al., which affected the performance of action recognition accuracy. With the advancement of technology, especially the application of depth sensors, Kinect can get human skeletal data directly, which makes the human action recognition based on skeletal data becoming a research hotspot again. Therefore, this paper studies human action recognition based on skeletal joints. The study is focused on human action representation and recognition, which based on skeletal data.First of all, this paper introduces the meaning of information entropy of human joint angle in human action representation, the relationship between information entropy, information quantity and variance of joint angle, the shortcomings of information entropy maximum algorithm based on skeletal joints. According to the problem of ignoring the noise in calculating the information quantity and low recognition rate of contrary process action, an action representation algorithm which fusion joint angle and movement trend information is proposed. The algorithm deals the noise problem using the principle of linear change of joint angle, sorts the information quantity and adds movement trend information. Finally, the action is indicated by skeletal joints with larger information quantity.Secondly, an algorithm combining 2D principal component analysis (2DPCA) and support vector machine (SVM) is proposed. This paper analyzes the action recognition algorithm of SVM. Although the efficiency of SVM is high when dealing with small sample, the way of classification is transforming data from low dimensional space to high dimension space. The amount of action data is larger, the efficiency of classification is lower. A reasonable feature extraction algorithm is required. Analyzing the shortcomings of PCA, the improved PCA deals with two-dimensional matrix directly, while keeping the relationship between matrix row and extracting meaningful action characteristic data. Finally, the action data is classified by SVM.Through a large number of experimental test on two datasets, it proves that the human action representation and recognition algorithms based on skeletal data in this paper have achieved good results, indicating that the proposed algorithms are feasible.