Abnormal Behavior Recognition In Power Generation

The intelligent grid is the leading direction of the current national grid development. The intelligence reflecte in various aspect of the production, including automatic security alarm. Safety is the premise of the electricity production and mostly relied solely on human monitoring so far. It always too slow to react to urgent mattrs. In order to reverse this situation, this article will engage in research on abnormal human behavior recognition in power generation on the means of computer vision.Key technologies in abnormal behavior, including the moving forground extraction, target identification, target tracking and behavior classification will be explored and studied here.Moving foreground extraction here consists of extracting image sequences, removing background and obtaining motion information of human goals and their belongings. Firstly, ther current frame is deal with symmetrical frame difference and background subtraction rspectively. Timing split result will be got by the or operator for the two methods.Then the timing forground region further processing by the improvement C-V model. The defects of traditional C-V model, such as slowly segmentation speed and empty boundary, are resolved by acceleration factor and weak boundary traction in homogeneity region. Late in the image pre-processing, the protential shadow noise is eliminated by HSV shadow model in Gaussian distribution.SIFT feature matching algorithm has a good performance for target rotation and zoom effects in template matching process, while its weak point is the real-time performance. In the stage of target recognition, descriptors are redefined. The more close the feature points, the higher weight assigned for the feature sub-vector. And a better searching strategy for feature vector is adopted. It reached the goal of optimization accuracy and real-time. Taking into account the color features of staff, such as helmets, overals, while influenced by light intensity effects, a tracking method with color and gradient histogram base on the particle filter is introduced here. Firstly, the three components of RGB color are quantified into different grayscale and normalized respectively. Then using orthogonal gradient descriptors construct gradient histogram in blocks. Gradient features make up the defect of color features in shape.An apporiate feature description method is extremely need in human behavior recognition. For the purpose of recognizing the abnormal behavior in power generation such as fall, fight and take off helmet, an idea of human outline geometric model is introduced here. The model includes human center of mass coordinates, the minimum bounding rectangle, bounding rectangle density, tilt angle, aspect ratio of rectangle, change rate of aspect ratio, distance between centroid and paticular object. Calculating model information by mathematical dedution and making a standard for abnormal behavior classification. Then constructing classifier by sample geometric information base on support vector machine and estimating the conduct actually occurred.