Study And Implementation Of Real-Time Arm Motion Recognition System

In recent years, the use of computer technology is getting more and more widely. And we need higher criterion and better environment to communicate with the computer, hoping that the whole process can be more nature, more convenience and can get the right feed back in a very short time. Traditional input device which are still being used today such as keyboard, mouse etc. limits the speed and naturality of Human-Computer Interaction (HCI). Whereas the new vision-based methods allow users to get rid of the dependence on equipment and bondage. They complete the interaction through catching and recognizing the body’s posture and motion. These methods typically use 2D cameras or depth cameras like Kinect to collect motion image sequence. Among them, the 2D camera is the most commonly used way for collection and we can also get many deep researches on this field. But most of them are not suitable for real-time environment for complicated detection algorithm and complex modeling method. In addition, the arm movement can be widely applied to game and robot interaction as an important interaction way. But there are not so many researches of arm motion recognition.Based on the above reasons, we propose and implement a real-time recognition system of the arm motion. The system works as follows:it captures the arm motion with a 2D camera. After gathering a certain amount of image sequence (typically a cycle of an arm movement), it detects the arm motion through three-frame difference algorithm. Then it models the arm with the two-dimensional stick model. And finally recognize the arm motion through the naive Bayesian classifier with the model. Generally speaking, the system is can recognize four arm motions for a single arm or two arms: waving, beckoning, circling and pushing forward. The system can recognize the motion in real-time. It can make out the correct recognition result in 4-6 seconds in normal computer. And a full motion usually lasts for 2 seconds, so the whole process time is just 2-4 seconds. In conclusion, the advantages of our system ar embodied in three aspects:firstly, it completes the recognition with only one monocular camera. So it’s more practical in the real world as it doesn’t need any other equipment and environment requirement. Secondly, the system adopts efficient motion detection algorithm and modeling method. So it can meet the real-time requirement with less computation and less time consuming. Finally at the classifying step, it counts the temporal information to the attribute value for training. And this helps to achieve more accurate classification result.