Research On Vision Human Action Recognition For Human-Computer Interaction And DSP Implementation

With the continuous development of artificial intelligence research and its application, the important research branch human-computer interaction technology was also gradually developed from the traditional contact to non-contact, from simple-interface to intelligent naturally, such as the voice recognition technology based on auditory and the image recognition technology based on vision. Compared with the voice signal, image signal is more of universality and stability, images can carry richer and larger amount of information than voice, so the human-computer interaction technology that based on vision attracted the interest to a large number of scholars and engineers, and it is a research hotspot currently. The key of Human-computer interaction technology that based on vision is human-centered image processing, especially the recognition and understanding of the body action posture. At present, this technology has been used widely or trying to use in service robot, intelligent household, wisdom and intelligent city security, and many other fields.The realization of the human-computer interaction technology that based on vision is generally only for limited simple body movements,such as running, walking and suddenly leaped, they all have strict recognition criteria, such as human body must be directly facing the camera, it cannot appear more than one person at the same time, it is only suitable for ideal simple background environment, etc. In addition, there are many shortages, such as the computational work and the memory consumption are so big, they are not suitable for embedded development of miniaturization, etc. To enhance the universality of human-computer interaction technology based on vision, strengthen its adaptability in real natural environment, this paper designed algorithm and developed deeper for the intelligence lives in the human-computer interaction technology based on vision,explored the recognition of the seven defined general human-computer interaction gestures in natural environment, deeply researched the key feature extraction of the defined action, and put forward a new method and thinking at the same time. Through experimental tests, the human body gestures recognition accuracy is above 79%, the efficiency of the interactive working is more than 75%, and it can adapt to different indoor environment, and can identify any space, direction and size independence.The key work and innovation in this paper are mainly the following:(1) The visual image processor is TI DM642, the master controller is ARM7, based on both of them, the paper designs and builds the intelligent home management system based on visual human-computer interaction experiment platform. This paper predefines a set of intelligent household interaction gesture, including some action commands, such as the choice of start, certain and negation, switch, functional flip up and down, pause, etc.(2) At stage of gestures target detection,this paper proposes a video background rapid extraction method that based on the nearest neighbor pixel gradient(N2PG) stability, and uses it in target detection of the human-computer interaction gesture recognition. At the same time, it solves the problem that some devices are difficult on target detection tracking in the natural environment, which the devices are based on visual embedded human-computer interaction.(3) In phase of gestures initial recognition, using the HMM model and based on the statistical characteristics of static characteristics form, solved the problem of poor anti-interference problem that using single frame image features to identify the action, and each of the static characteristics in the form using normalized geometric moment and low frequency Fourier descriptor to extract feature, it can identify any space, direction, size and rotation independence.(4) At the second check stage of gestures recognition,this paper proposes an eight- two code(8-bit binary code) calibration method, which is based on the speed direction variation characteristics, realizes the fast matching verify or supplement recognition of the preliminary identification, improves greatly of the gesture recognition accuracy and robustness of human-computer interaction system.