Development of Object Detection-based Surveillance System for Campus Night Patrol Robot

Abstract

Campus security at night faces challenges from manual patrols limited by human fatigue and poor visibility, creating gaps in monitoring theft, vandalism, and other threats. This study develops an object detection-based surveillance system for a campus night patrol robot that integrates low-light image enhancement and real-time detection. Using the ExDark dataset refined to five classes (people, cars, motorbikes, bicycles, buses) with YOLO-format annotations, the system applies 640×640 resizing and Zero-DCE enhancement to improve contrast and visibility without artifacts. The YOLOv11-S model was fine-tuned via transfer learning from COCO weights using ADAMW optimization, cosine annealing (learning rate 0.001), and mosaic/mixup augmentations. On a GPU-equipped laptop it achieved [email protected] of 0.863, [email protected]:0.95 of 0.647, precision of 0.892, and recall of 0.831, delivering 20–30 FPS on Jetson AGX Xavier with Logitech Brio 4K input. Results indicate strong vehicle precision but class-imbalance bias toward human misdetections; ablation studies confirm Zero-DCE boosts mAP by ~15%. The main contribution is a lightweight, edge-deployable framework combining Zero-DCE low-light enhancement with fine-tuned YOLOv11-S for real-time, high-accuracy nighttime surveillance. This framework outperforms literature baselines (mAP ~0.5) and offers a scalable solution for automated campus security, with future work focusing on TensorRT optimization and dataset balancing.