Integrated Autonomous Exploration and Deep Learning Vision for UAV Search and Rescue Missions

Unmanned Aerial Vehicles are crucial for Search and Rescue missions, yet achieving full autonomy requires solving two critical challenges within strin gent edge-hardware constraints: robust collision-free navigation and accu rate perception of small-scale targets. This thesis proposes a modular frame work addressing both domains. First, we develop an autonomous exploration pipeline integrating an A* global planner with an Improved Artificial Poten tial Field local controller. Rigorous statistical analysis of simulated cam paigns demonstrates that the system maintains highly stable flight velocities and sub-millisecond planning times, guaranteeing safe navigation regardless of environmental clutter. Second, we introduce an optimized human detec tion model based on YOLO11s. To detect minute, camouflaged targets, we implement a two-stage transfer learning pipeline alongside targeted architec tural modifications: a Parallel Bi-directional Feature Pyramid Network and a high-resolution P2 detection head. The model achieves a peak mAP@50 of 0.814 with a computational footprint (28.8 GFLOPs) viable for standard edge accelerators. However, executing both dense visual recognition and complex control concurrently exposes the bottlenecks of traditional von Neumann ar chitectures. The thesis concludes that achieving true micro-UAV autonomy necessitates a paradigm shift toward Neuromorphic Computing and Spiking Neural Networks to unlock ultra-low latency and energy efficiency.