Multi-view Human Behavior Recognition And Estimation

Human behavior recognition based on computer vision has been a key point of human body analysis. It has been applied in camera monitoring system, intelligent furniture system, sports analysis and so on. In a word, human behavior recognition has a bright future for science research. As the cameras have been used in all kinds of place, such as banks, park places and shops, it alarms only the dangerous occasions happened. The alarm systems are not able to alarm if those begin. According to it, we could eliminate the loss as much as possible if we have an auto-recognition system which could inform people when illegal behavior happens. What’s more, human behavior recognition provides a all new interactive mode with more information. In a word, human behavior recognition indeed makes sense.As we can see, many of them concentrate on behavior recognition with single camera. Since of that, it has obvious disadvantages. First of all, the method is sensitive with illumination which leads to poor result. Second, the shelter. Because of the stationary view, we have to suffer from sheltering sometimes which affect the segmentation or other steps. Lastly, poor accuracy. It will lead to a poor accuracy of recognition with bad behavior feature. In order to overcome those elements, I am going to have a human behavior recognition algorithm with multi-camera. Depending on two different view cameras, it is possible to solve the problems caused by illumination, shelter, angle of view and so on.Based on single view human behavior recognition algorithm, We compare the GMM with background subtraction. With the excellent result, background subtraction remains the better idea for image segmentation. Then, by the advantage of k-means algorithm, We are going to combine those key-frames to the human behavior template library which is extract by pervious step. The human behavior template library always includes three types:walking, squatting and waving hands. At the last step, we used Borda voting algorithm as well as Newton’s law of cooling to decide the biggest probability status as the result within all behaviors by two different views. With a lot of experiments in complicated environment, the human behavior recognition algorithm shows a excellent performance in segmentation, capture the Mu moment and high accuracy. Additional, we have applied the algorithm in ATM security system to protect people from crime.