Design And Implementation Of Real-Time Embedded System Of Moving Human Detection

To detect and recognize the moving object in image sequences is an important research direction in the current area of computer vision and pattern recognition, it also has an important standing in the area of intelligent visual surveillance, image searching and advanced human-computer interaction. In this paper, we study effectively the relevant contents of pedestrian detection in image sequences, and accomplish a real-time system of moving human detection based on embedded platform.The detection of moving human can be divided into two parts:to segment the moving object in image sequences and to classify the extracted moving target. In the circumstance that camera was fixed, the most common algorithm of detecting moving object in image sequences is the method of background subtraction, which always need to construct the background model and update the background image timely, and it's easily influenced by the change of light. The method of optical flow can extract the whole moving object without any information of background, but the operation is complicated, bad real-timing and need the support of special hardwares. The method of temporal difference has highly adaptivity and real-timing, however, which exists the problem that the overlap of object between two adjacent frames is hard to detect, and the inside of object is easy to cause cavities. In view of the feature that resource limited of embedded platform, we have made an improvement based on the method of three-frame-differencing, exclusive disjunction is adopted to replace the final operation of conjunction, which carries out the accumulation of motions, so that we can get more areas of moving object and reduce the inner cavities of object. The method of maximum between class variance is used to segmentation the gray image by adaptive threshold, and then the method of mathematical morphology is adopted to dispose the foreground object, which can reduce the inner cavities of object further more. Finally, the method of extraction of connected domain is used to extract the moving object.The human recognition and behavior understanding are always recognized as the hot research point of computer vision, most human recognition methods are based on the knowledge of human structure, skin color and movement, usually we choose the feature vectors based on human structure, and then use the method of pattern classification for prediction. At present, there are many pattern classification algorithms, mainly including bayes classifier, support vector machine, K nearest neighbor method and so on. Judging the extracted object whether belongs to human or non-human is a dichotomous classification problem, considering the real-timing of embedded platform, we use bayes classifier as the classification method in this paper. In the choice of features, firstly according to the characters of human structure we separate the sample images into pieces, secondly we statistic the grey values of all pixels in each piece, finally we gather all the statistics from the components of vertical and horizontal and combine them as the feature vectors of the classifier.At last, we apply the algorithms of moving object detection and human recognition to embedded video surveillance, which can make an intelligent judgement whether detected a single moving human or not in the static monitored scene without any human intervension, when somebody was detected the system would automatically trigger an alarm and save all the captured images at that time into the assigned location of the system. The experiment results show that it can detect the moving human in static monitored scene rapidly and effectively, moreover, it's robust to the normal change of light.