Light Weight Real-Time Burglary Detection and Inspection in Low Light Surveillance Videos

Crime is a serious concern affecting lives, businesses, and livelihood globally. Crime detection using surveillance video content has been an active area of AI research for a decade with datasets and methodologies evolving to detect and classify anomalous activities. However, there isn’t any active research on improving anomaly or crime detection for surveillance videos shot in low light conditions despite significant advancements achieved in dark or low illumination image and video enhancements. Recent studies have proposed using cameras with night vision capability or other equipment to overcome the problem of detecting intruders in low illumination scenarios. However, this requires additional hardware setup and does not address the problem of existing surveillance cameras without night vision capability. There also isn’t any active research to advance beyond just detection and/or classification to inspect the details of the crime. Moreover, the state-of-the-art methodologies are compute-intensive and cannot be deployed on an edge device for real-time crime detection. The proposed research addresses these problems by implementing a lightweight AI pipeline consisting of high speed low compute image enhancement network to enhance the low light videos, and perform real-time burglary detection on a Raspberry Pi 4 edge device using Dual Input 2-dimensional Convolutional Neural Network (C2D). On detecting a burglary event, the pipeline further triggers a YOLOv5s based real-time Burglary Inspection System on the edge device to report the number of burglars. Using Burglary as the category for this study, the research further advances anomaly detection in low lighting conditions as well as paves the way to introduce and explore inspection in burglary and other categories of crime or anomaly as future work. Since a benchmark crime or anomaly detection video dataset shot in low-light conditions is not available, the research additionally synthesizes low-light videos from available videos shot in all lighting conditions and proposes a benchmark low-light crime detection dataset to help advance research in this area.