Human Action Recognition And Suspected Cheating Behavior Detection Based On Latent SVM

Human action recognition based on vision can be applied to the video surveillance system,human-computer interaction, intelligent robots, intelligent home systems, video analysis, andelectronic entertainment, thus make it one of the most active field and an important part ofcomputer vision. The mainly research of action recognition is in video, and the one from stillimage has not been widely concerned. Compared with the video the information about theaction still images contain is not rich enough, which causes the issue more difficult. Butwhether the properties of the action itself or the current research results show that therecognition from still images in certain areas is feasible and reliable. This article based on thecurrent advanced theories and techniques of pattern recognition makes summary and analysison some recent literatures, and then proposes a simple and effective action representationconsidering the practical applications, and finally achieves the training and testing for themodel whose structure is designed corresponding to the representation. According to thecharacteristics of abnormal behavior in the examination room we designed an efficientalgorithm for applying action recognition based on still image to the detection in the videoframe sequence, and the testing in actual video proved that the action recognition based onstatic images could be used in video analysis, which has a positive significance for theintelligentization of video surveillance.Firstly, this paper created an independent dataset of human action. We collected sevencommon daily actions as research subjects including reading, exam, playing violin, playingguitar, walking, and running. These pictures, with about200for each class and total of1500including3000objects, are collected on the internet via entering the keyword such as"running" to the web search engine. We specifically developed an efficient tool,AnnotationV2.0, to annotate the action among800JPG pictures containing1200objects andgenerate annotation information recorded in800xml files. This work provides a sample setfor the model training.Secondly we designed multi-viewpoint action model using part-based models anddeformable template. The design of the action model is a multi-viewpoint mixed templatewith4kinds of viewpoint: front, left/right45°, left/right side and back. This taxonomy canrespond to significant appearance variation of action in static images caused by differentcamera viewpoint. In order to improve the accuracy of the model in detecting the personperforming action under complex background, we use the HOG feature extraction method toconstruct the templates. To make the templates be able to succinctly express the content of theact, we propose the concept of Action Coverage Area (ACA), that is, a complete description of body parts and items involved in the behavior. The ACA will not describe any of thenon-coverage area. Meanwhile, in order to make the model has a great robustness in the caseof incomplete target we add a component called Action Core (AC) to the ACA. AC has asmaller area thus with a high speed of detection and can be used to assist in the analysis ofvideo content. Taking into account the non-rigid changes of the human actions, namely thenon-rigid changes of the body parts and action extent, we used part-based model anddeformable templates to design the local parts of the ACA and AC which are in the form ofHOG features. Experiments has showed that model with these configurations is able to adaptto intra-class variation.Thirdly, we studied the classifier Latent SVM and use it to train the action models whichachieved a high performance. Latent SVM introduces latent variables on the basis of thelinear SVM. During training the action model with Latent SVM we treated the viewpoint andthe spatial constrains of the parts as the latent variables. In the process of training the LatentSVM provides a high efficient training set data mining method which speeds up theconvergence of the training. At the same time, the efficient object detection algorithmcontained in Latent SVM as a reliable technology foundation makes it possible to apply theaction model to the video analysis. The experiments have proved that use the Latent SVM asthe action model classifier is appropriate.Finally, we used the action models which were trained from still image to design thealgorithm for detecting abnormal behavior in examination room. Intelligence is one of thetrends of the development of video surveillance technology. In the domestic, videosurveillance is more widely used in the examination room, but the degree of intelligence is nothigh. At present, the method of information filtering, retrieval and storage will become thebottleneck of the application development. How to monitor the video efficiently and targetedto handle the vast amounts of video data became the current problems to be solved. In thisregard, we designed an efficient algorithm carrying out the detection of the abnormal behaviorduring examination on the basis of the sufficient statistics and analysis of the examinationroom scene. The algorithm includes the detailed design of the examinees informationstructure and content, in which we proposed the concept of Three-Dimensional Attention(TDA). The information structures are firstly well to explore the historical relationshipbetween video inter frame which improve the detection efficiency, and secondly provide thestatistical data for assessing the examination room condition and personal integrity.