Human Action Recognition Based On Image Feature And Skeleton Feature Of Kinect

Human action recognition is a critical research direction in computer vision research field. The study and analysis of human action recognition involved in many fields such as computer vision, artificial intelligence, have an important research value. The traditional methods for on human action recognition based on RGB video sequence, these methods not only have high computational complexity, but also are sensitive for the change of illumination, background and the angle. Some depth cameras we used in the past, limit their application by the expensive price and huge volume. Kinect is favored by the researchers because of the good cost performance.In this paper, the related problem with the human action recognition based on Kinect were studied in detail, including the image preprocessing, depth feature extraction, skeleton feature extraction, key frame extraction and SVM classification. We propose a human action recognition method based on key frames and fusion more features. The main works are as follows:A method was proposed to extract the feature for human action.. Firstly, the image is pretreated, such as binarization, denoising. Aspect ratio and motion trail are extracted. These two features can represent the changes respectively of the human body in the plane and space from a motion video sequence. At the same time, human skeleton joints position and skeleton angle are extracted. These two features can represent the changes of relative position between the human body from a motion video sequence.In this paper, we propose a human action recognition method based on key frames and fusion more features in human action recognition process. First, we use dynamic time warping algorithm to handle the temporal misalignment. Second, the k-means algorithm is used to extract sample center, and then the sample center is used to extract key frames from the video sequences, and then extracting the depth features from the original video sequence and extracting the skeleton features from key frames. And two kinds of features are fused. Finally, the feature of fusion is used by SVM classifier for classification and identification. The way of extracting key frames can not only cut down the number of the feature data used for human action recognition, but alse have a good effect on the recognition rate due to eliminating the redundant data, but also reduce the computational complexity and the time spent on action recognition. As a result, the accuracy and timeliness of the human action recognition can be improved.We evaluate the performance of the human action recognition on the public depth datasets and a custom data set. Compared with the other methods, from the experimental results, not only the recognition accuracy but also the real-time capability has been improved by using the human action recognition based key frames. In conclusion, the effectiveness of the method proposed in this paper is proved.