BUnmanned aerial vehicles (UAVs) have shown great potential in patrol

surveillance

and search and rescue applications due to their flexibility and efficiency. This paper proposes a vision-based obstacle avoidance system for UAV patrols. This system integrates a high-resolution RGB camera and a dynamic vision sensor (DVS). It utilizes a lightweight YOLOv5 model for obstacle detection and an optimized dynamic window approach (DWA) for real-time path planning. Through multi-sensor fusion

threat level assessment

and dynamic speed adjustment

the system significantly improves robustness in complex lighting

weather

and occlusion conditions. In ROS and Simulink simulations

as well as field tests on campuses and industrial parks

the system demonstrates impressive performance: obstacle detection accuracy exceeds 95% under variable conditions

path planning latency is less than 100 milliseconds

and obstacle avoidance success rates reach 95% in low-density scenarios and 90% in high-density scenarios. Compared to the deep reinforcement learning algorithm D3QN

the system achieves a 13.3% reduction in path length and a 20% improvement in success rate. Deployed on low-cost Raspberry Pi 4 hardware

the system achieves 20 FPS processing with less than 5W power consumption

providing a practical solution for scenarios such as border patrol

industrial monitoring

and urban surveillance. Innovations include a lightweight convolutional neural network (CNN) architecture

adaptive obstacle avoidance strategies

and a modular software framework

addressing the challenges of accurate recognition

rapid planning

and coordinated optimization of patrol efficiency in complex environments.