Advanced Image Processing Techniques for Wildlife Monitoring and Conservation

Abstract

Rapid, non-invasive monitoring is critical for safeguarding wildlife in both African savannas and fragmented European forests. This thesis designs and evaluates an edge-optimised computer-vision pipeline that combines classical motion detection with a MobileNetV2-based object detector to identify elephants, rhinos, buffaloes, and zebras in camera-trap imagery while also flagging human intruders. Using a curated, multi-habitat dataset (30 000 Kenyan and 15 000 Hungarian images) we achieve 0.91 overall accuracy and per-class F1 scores up to 0.97, all on a Jetson-Nano power budget of <5 W. A confusion-matrix-driven error analysis highlights occlusion and class-imbalance as the primary failure modes and motivates targeted data-augmentation and lightweight attention modules. Finally, we present a bill-of-materials and ethics-compliance workflow demonstrating how the system can be deployed for real-time anti-poaching alerts and long-term biodiversity surveys. The framework is fully open-source and adaptable to new regions with minimal retraining.