Traditional unmanned aerial vehicle (UAV) lighting systems exhibit poor adaptability in complex environments, high energy consumption, and uneven coverage. This paper proposes an autonomous planning and optimization method for UAV lighting tasks based on multi-sensor data fusion. The method establishes a three-tier fusion mechanism comprising “feature layer-state layer-decision layer”: at the feature layer, a dual-stream convolutional neural network (CNN) extracts multimodal features from visual and depth images; At the state layer, an improved unscented Kalman filter (UKF) fuses IMU and visual odometry data to dynamically estimate UAV position, velocity, and attitude, incorporating an adaptive noise adjustment mechanism to enhance environmental adaptability; at the decision layer, D-S evidence theory integrates multi-sensor lighting demand assessments to resolve information conflicts and achieve robust decision-making. Furthermore, a weighted objective function incorporating coverage efficiency, energy consumption, and path smoothness is designed. An improved particle swarm optimization (PSO) algorithm is proposed for offline initial path generation, combined with Q-learning for online dynamic adjustment of weight parameters, enabling adaptive optimization during task execution. Experiments conducted on the DJI Matrice 350 RTK platform across three typical scenarios demonstrate that the proposed method reduces position estimation error by 73.5%, achieves 98.2% illumination coverage, and lowers energy consumption by 25.8% in complex obstacle environments—significantly outperforming conventional approaches. This method effectively enhances UAV illumination task execution capabilities and autonomous intelligence in complex dynamic environments.
Guo et al. (Sun,) studied this question.