Research Of ATM Compartment Tailgate Detection Algorithm

The trailing detection of the ATM compartment is a specific application of the intelligent video surveillance system, and which is the mode of making the postmortem analysis into the pre-warning before things happening and the alarming when things happening. This paper summarizes the current status of research on the trailing detection techniques, and analyzes several difficult factors about the accuracy of the trailing detection, the trailing detection algorithm of ATM compartment is designed. The feasibility and accuracy of the algorithm is validated by test. The main work of this paper is as follows:1. The threshold image is the average of all difference images, which improves the fixed threshold of the Surendra background update algorithm.The noise of the image can be reduced to some extent based on the theory of calculating the average of the image. First we can get all difference image of a video sequence using the method of two adjacent frame difference,then get the average of all difference images by the calculation method proposed in this paper.Finally the average image is the threshold image.2. The head and shoulders of human is detected by using local gradient pattern features and binary histograms of oriented gradients features.This paper proposes the two novel local transform features to detect the human head and shoulders. One is the local gradient pattern, and which does not directly use the gray value information of the neighborhood pixels, but gradient information, so which can better describe the texture information. Another is the binary histograms of oriented gradients, and the BHOG feature represents each block using the 8bits,which is suitable for real-time processing because the number of dimensions reduces greatly.3. The feature confidence decides which feature to choose classifier for target recognition.SVM the local gradient pattern features and binary gradient direction histogram features of thesample is learned and trained by SVM, then two feature classifiers is obtained, finally to choose which character classifier by means of the method of the confidence of feature for target recognition. My method is feasible by testing in this paper.