VIOLENCE DETECTION IN CROWD FOOTAGE: ENGINEERING STATISTICAL FEATURES USING TRANSFORMED OPTICAL FLOW
Video surveillance technology is becoming increasingly common and is often used to increase safety in public spaces. However, the effectiveness of the video information on its own is questionable as it relies on someone manually reviewing the footage, in real time or retroactively. The current research intends to improve video surveillance technology by using computer vision and machine learning techniques to automatically detect a violent event within a crowded scene, in real time. Meaningful information is extracted from the raw, gray-scale pixel data. This is done through optical flow feature extraction, then a projection of those values onto a plane which approximates the ground in the reality of the scene. The projection is done as a means to account for perspective distortion affecting the optical flow calculation. The dimension of the feature set is reduced through quantizing the frame and following several statistics through time. Random forest variable ranking is leveraged to further reduce the feature set. Promising results are found using simple cutoff classifiers in the target region of the frame.