Efficient Video Anomaly Detection using One-Class Support Vector Machine

Nowadays, video anomaly detection is crucial task in various applications and the increasing amount of video data generated from various sources has made it essential to develop efficient anomaly detection methods. However, the existing Local Outlier Factor (LOF) faced challenges in detecting anomalies due to the complexity and variability of video data. Hence, a One-Class Support Vector Machine (OCSVM) is proposed to identify unusual events in videos by separating normal and anomalous data through a hyperplane. Initially, the input video data is collected from University of California, San Diego (UCSD) pedestrian dataset which consists of grayscale video footage captured along walkways at the UCSD. Then, shot segmentation using boundary detection algorithm is done to divide the video into meaningful shots. Then, Vision Transformer (ViT) is employed to identify the relevant characteristics within the frames. After that, key frame extraction process selects the representative frames from video sequence by capturing significant changes or events. Finally, the proposed OCSVM is introduced to detect unusual patterns by learning a boundary around normal data points. From the results, the proposed OCSVM achieved better results when compared to existing Convolutional Neural Network (CNN) in terms of accuracy (95.74%), precision (92.16%), recall (92.11%), and F1-score (93.27%) respectively.