AIGuard: Anomaly Detection in Surveillance Videos with YOLOv8

Abstract

This study introduces AIGuard, an AI-powered system designed to detect abnormal behaviors in public surveillance footage. Built on the YOLOv8 object detection framework, the system identifies specific anomalies-such as individuals riding bicycles or skateboards in pedestrian zones, or leaving behind unattended objects-using the UCSD Ped1 dataset. This dataset is widely used in anomaly detection research but poses significant challenges due to its grayscale imagery, low resolution (158 × 238), high-angle camera views, and varying environmental conditions such as low lighting and rain. To overcome these visual limitations, we propose a multistage image enhancement pipeline that preprocesses video frames before detection. This pipeline combines three specialized neural networks: LLNet for low-light enhancement, DehazeNet for rain and haze removal, and FSRNet for super-resolution upscaling. By improving the visual quality of the input frames, the system enhances object visibility and overall detection reliabil-ity-especially in difficult surveillance scenarios. The anomaly detection itself is based on a transparent, rulebased logic that classifies behaviors as normal or abnormal based on object interactions and spatial relationships. Detected anomalies are clearly visualized in real time with bounding boxes and textual alerts embedded directly in the video stream. Despite the limitations of the dataset, our system demonstrates improved accuracy in identifying and describing abnormal events, largely due to the enhanced image clarity and resolution provided by the preprocessing stage. These results highlight the practical potential of combining lightweight detection models with interpretable logic and intelligent image enhancement. The proposed pipeline offers a robust, low-cost solution for real-world surveillance scenarios-especially where full semantic scene understanding is not feasible. This enhancement framework addresses common environmental challenges that are often overlooked in conventional anomaly detection approaches.