Intelligent Surveillance Framework for Physical Abuse Detection

Abstract

Anger has been an important human emotion since time immemorial. Showing anger newlinein physical forms have been a common trait among certain people irrespective of caste, creed, newlinegender, education or financial background. Physically causing harm to the people around them, newlineespecially weaker ones like children, women and older people can cause severe bodily damage newlineand mental trauma for the victim. Surveys and statistics from renowned organizations from newlinedeveloped countries like the US and developing countries like India show startling data about newlinewomen and children getting physically abused. Campus violence is also a social evil. Schools, newlinecolleges, child care centers, old age homes, hostels are common places prone to have physical newlineabuse and violence. Monitoring these places using CCTV cameras also requires human newlineintervention for its effective detection. Hence intelligent surveillance systems are the need of newlinethe hour. newlineHuman action recognition is a complex task and an active area of research in the field newlineof Computer Vision and Machine Learning. Image processing and computer vision algorithms newlinehave limitations in expressing complex human action in a scene. Human action recognition is newlinenot just the movement pattern of human body parts, the context, culture, the collective newlinebehaviour of all persons in the scene matters. Hence feature extraction is the complicated part. newlineDeep Learning is the ideal technique for handling this complexity. Further the advance in GPU newlinearchitecture and availability of huge amounts of video data due to rise in social media usage, newlinehas given Deep Learning an edge over other technologies. newlineIn the first stage of research work, the importance of the human hand in the scene was newlinerecognized and focus on capturing its movement as the most important factor defining the newlineaction. Computer vision algorithms were used for hand feature extraction, and classification newlinewas done using DL model. Though the results were good, the complexity of the scene was newlinelimited by the feature extraction algorithm.