A Unified Video Summarization for Video Anomalies Through Deep Learning

Abstract

Over the last ten years, integrated video surveillance systems have become increasingly important in protecting public safety. Because a single surveillance camera continuously collects events in a specific field of view at all times of day and night, a system that can create a summary that concisely captures key elements of the incoming frames is required. To be more specific, due to time constraints, the enormous amount of video footage cannot be properly examined for analysis. As a result, it is vital to compile a summary of what happened on the scene and look for anomalous events in the footage. A unified approach for detecting and summarizing anomalous events is proposed. To detect the event and compute the anomaly scores, a 3D deep learning approach is used. Afterward, the scores are utilized to visualize and localize the anomalous regions. Finally, the blob analysis technique is used to extract the anomalous regions. To verify the results, quantitative and qualitative evaluations are provided. Experiments indicate that the proposed summarizing method keeps crucial information while producing competitive results. More qualitative results can be found through our project channel: https://youtu.be/eMPMjiGlCQI