A Human Action Recognition Method Based On Computer Vision

At present,human action recognition technology has become a very challenging direction in the field of computer vision.The biggest difference between action recognition in video and action recognition in images is that video not only contains information on timing,but also requires a huge amount of computational support.Moreover,it is also affected by many external factors,such as the video dynamic background is mostly,the camera angle is changed,and the illumination changes.With the rapid development of science and technology and the enhancement of chip computing capabilities,human action recognition technology based on computer vision is now receiving more and more attention from researchers.In this thesis,the performance of Naive Bayes and Semi-Naive Bayes classification algorithm is compared on image recognition,and extract the features of the image by Gaussian blur,Gray-scale processing and Image binarization.The classifier performs recognition and comparison.The experimental results show that the Semi-Naive Bayes has better than the Naive Bayes algorithm in image recognition.Subsequently,this thesis proposes a method based on the combination of DT and Semi-NBC.The basic idea is to use the optical flow to extract the human motion trajectory information in the video,and then extract some feature descriptors such as HOG,HOF,and MBH from the trajectory information.Where HOG describes static appearance information in the video,HOF describes local motion information in the video,and MBH is a gradient value for calculating the optical flow.This thesis uses Fisher Vector coding for the extracted features and use the Semi-Naive Bayes classifier to train the recognition classification based on the encoded results.The experimental results show that the proposed method based on the combination of DT and Semi-NBC improves the recognition accuracy in the video recognition task compared with the DT algorithm.