An Intelligent Algorithm For Detecting Cheating Behavior At The Exam Site

Intelligent behavior analysis is an increasingly important research topic in computer vision. Due to the uncertainty of characteristics of the external environment and the behavior of target objects, it is difficult to present a general method for the analysis of the intelligent behavior. Nevertheless, for specific scenarios and characteristics, it is possible to design an algorithm with strong robustness and precise accuracy.Intelligent Vision Surveillance (IVS) is a typical application of intelligent behavior analysis technology. The work presented in this thesis researches intelligent analysis method for detecting cheating behavior during examinations, which is a particular aspect of the application of IVS. With respect to characteristics of the environment of examination rooms and the candidates'behavior, this thesis presents a set of intelligent analysis methods to tackle with suspicious cheating behavior in examination. Experiments show that the method is robust with low time-consuming. Our contributions can be mainly categorized into following parts:(1) To achieve the purpose of this study, this thesis designed an intelligent analysis process for detecting suspicious cheating behavior. It is from the classic designation of intelligent analysis process, which includes desks detection, people counting and behavior analysis.(2) This work proposed a method for desk detection in the standard college entrance examination environment, which used Perspective Transformation, gradient template matching, and Random Sample Consensus (RANSAC) to correct the primary detection results. In comparison with RANSAC, the Least Squares algorithm was proved that there exist the limitations for image anti-noising.(3) This work introduced a method for students counting, which is based on camera calibration and similarity matching. By comparing Normalized Cross Correlation Algorithm with the classic background reduction method and Histograms of Oriented Gradients (HOG), it proves that Normalized Cross Correlation Algorithm is superior in identifying students in low-resolution images. To reduce incorrect detection results leading by the occlusion during different students, a method based on sequence matching and filling background was proposed.(4) This work employed an analysis model according to motion blobs, which divides cheating behavior into four categories. It uses the background updating algorithm to extract motion blobs of students and designs particular algorithm with respect to particular cheating behavior. As is based on motion blobs, the analysis model effectively reduces the flawless caused by the uncertainty of the color, texture and other characteristics of the environment, which increase the robustness of the system. The analysis algorithm mainly uses geometry features of motion blobs, hence the mathematical model is simple constructed with low time-cost.