Research on Target Recognition and Path Optimization for Unmanned Aerial Vehicles Based on Multi-Sensor Fusion

Abstract

Addressing challenges in UAV ground observation under complex environments—such as variable target scales, disordered distributions, and limitations of single-sensor perception—this study integrates multi-sensor fusion for UAV target recognition and path optimization. First, to enhance the robustness and accuracy of target detection, an improved YOLO detection model is proposed. Utilizing Darknet-19 as the backbone feature extraction network, this model removes fully connected layers, incorporates prior anchor box mechanisms, and optimizes training strategies. These modifications significantly improve the model's adaptability to multi-scale, irregularly distributed targets. Using a UAV equipped with visible light and infrared thermal imaging sensors for aerial photography, a multimodal dataset was constructed to train the network. This achieved deep fusion at the visible light and infrared feature levels, significantly improving target recognition accuracy and recall in complex lighting and occlusion scenarios. Furthermore, to achieve efficient task execution, this study designed a real-time path optimization algorithm that tightly couples perceptual information. This algorithm uses multi-sensor fusion recognition results (including target location, category, and confidence level) as key inputs to establish a multi-objective optimization function encompassing mission completion time, detection benefits, and flight risks. By introducing an improved Model Predictive Control (MPC) framework, it dynamically plans the UAV's observation pose and flight trajectory, enabling autonomous balancing between exploration and exploitation. This ultimately generates an optimal or suboptimal path that maximizes target search efficiency while ensuring flight safety and energy efficiency.