AI-DRIVEN WILDLIFE PROTECTION: REAL-TIME DETECTION, CLASSIFICATION, AND ALERTS

Abstract

Camera traps have become a core tool for wildlife monitoring, but the large volume of images they produce creates significant delays between event capture and analysis, particularly in remote conservation areas with limited connectivity. This work proposes an edge-based, AI-driven wildlife protection system that performs real-time detection, false-trigger elimination, and multi-channel alerting directly on low-cost embedded devices. An optimized YOLOv8 object detector is deployed on Raspberry Pi 4 and NVIDIA Jetson Nano platforms to localize and classify wildlife species from camera-trap images. A multi-stage false-trigger pipeline combining motion analysis, confidence thresholding, size filtering, and temporal consistency is integrated to suppress empty and non-wildlife frames, while a store-and-forward mechanism buffers high-priority events locally and synchronizes data with the cloud when connectivity becomes available. Experiments conducted on 15,000 labeled images spanning 12 wildlife species achieve a mean average precision of 0.84 at IoU 0.5, with precision 0.89, recall 0.86, and F1-score 0.87. The optimized models yield average inference latencies of approximately 152 ms on Raspberry Pi 4 and 78 ms on Jetson Nano, enabling near real-time operation, while the false-trigger elimination pipeline reduces data overload by 72% and the end-to-end alerting latency to 3.2 s. These results demonstrate that the proposed architecture can provide accurate, low-latency wildlife