Watchful Eye Recognition of Abnormal Human Activities in Wild Video

Abnormal human activity detection is crucial for security and public safety, but traditional Machine Learning (ML) techniques face challenges. A major limitation is the reliance on manual feature extraction from videos or images, which can impact classification accuracy due to its dependence on human factors. To overcome these challenges, this paper introduces Watchful Eye, an innovative approach that combines the Lucas-Kanade Optical Flow algorithm, Canny Edge Detection, and a Convolutional Neural Network (CNN) to automatically capture and analyze aggressive frames from video footage. Unlike conventional methods, Watchful Eye leverages optical flow to detect rapid and unusual movements and employs edge detection to extract critical structural information, which is then fed into a CNN for classification. The proposed CNN model is trained using the UCF Crime dataset. The accuracy and the AUC score achieved with only 25 iterations of training were 72.03 percent and 83.75 percent, respectively.