Research And Implementation On Object Classification Oriented To Action Recognition

In recent years, with the development of computer vision and image processing, and the improvement of CPU/GPU and other related hardware performance, more and more computer vision technology has been used in industrial production and real life. As people’s living standards being improved continuously, more and more people have a strong interest in sports, and tend to spend a lot of time and money to the campaign. So it is in dire need that that there is a motion analysis equipment or software for improving their skills. Therefore, motion analysis applications based on computer vision technology get a lot of favors by people of all ages. However, there is only some simple automatic functions in a number of vision-based motion analysis software on the market, and they cannot analysis and process user’s actions behavior automatically. To this end, we propose a new method of object classification for action recognition in the video stream, which can recognize and classify human motion in video stream in real time, and can automatically identify whether there are particular behavior action occurs or not. Specific contents of this thesis are:1) A latest detection method to human body detection and action classification was applied to our approach, by improvement of the original method, and effectively improve the classification accuracy, especially the human body detection accuracy can be up to 97% in natural scenes on our own dataset in wild scene. Also, the human position can be located more precisely, and the rate of overlap of detection results and the actual position of the body is more than 95%.2) Based on the research of human motion detection and classification, we proposed a new method of object classification oriented on action recognition in the video stream. By classifying the key action in continuous video stream, an observation sequence can be obtained in real time, then a specific action can be recognized. So as to achieve that identifying a particular behavior from a continuous stream in real time.In order to verify the effectiveness of the proposed method in practical applications, we also implemented a real-time motion video recognition system, and by using our method, the speed can reach to 30fps on iPhone 5s, and the accuracy is more than 90% on our own dataset.