Deep Reinforcement Learning-Based Target Detection and Autonomous Obstacle Avoidance Control for UAV

To address the challenges faced by distribution network monitoring systems—such as significant variations in anomaly scale, frequent missed and false detections of small-scale faults, and the need for real-time operational control—this paper proposes a lightweight multi-scale feature fusion detection network combined with a deep reinforcement learning-based autonomous control strategy, forming an end-to-end intelligent perception and decision-making system for distribution networks. To enhance detection accuracy and computational efficiency, a lightweight feature fusion network (Grid_RepGFPN) is designed, and a novel feature fusion module (DBB_GELAN) is proposed, which significantly reduces model parameters and computational cost while improving detection performance. Additionally, a feature extraction module (FTA_C2f) is constructed using partial convolution (PConv) and triplet attention mechanisms, combined with the ADown downsampling structure to improve the model’s capability to capture spatial and electrical measurement details. The programmable gradient information (PGI) strategy of YOLOv9 is further optimized by introducing a context-guided reversible architecture and a Grid_PGI method with additional detection heads, thereby enhancing deep supervision stability and reducing semantic information loss. Based on the detection model, a real-time operational control strategy is developed using deep reinforcement learning, enabling autonomous fault response, load adjustment, and network optimization through a state–action–feedback optimization loop. Experimental results on multiple distribution network simulation platforms demonstrate that the proposed LMGrid-YOLOv8 model outperforms YOLOv8s, with improvements of 4.2%, 3.9%, 5.1%, and 3.0% in precision, recall, mAP@0.5, and mAP@0.5:0.95, respectively, while reducing parameters by 63.9% and increasing computation by only 0.4 GFLOPs, achieving a favorable balance between performance and resource consumption. Inference experiments on edge computing platforms confirm that the proposed model maintains high detection accuracy under real-time constraints, demonstrating strong applicability to real-time distribution network monitoring. Furthermore, class activation map-based visual analysis reveals the model’s superior capabilities in detecting small-scale faults and processing high-resolution network measurement regions.

@article{Zhao2025,
  title     = {Deep Reinforcement Learning-Based Target Detection and Autonomous Obstacle Avoidance Control for UAV},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {9},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Like Zhao and Hao Liu and Guangmin Gu and Fei Wan and Yanyang Feng},
  doi       = {10.14569/IJACSA.2025.0160908},
  url       = {https://doi.org/10.14569/IJACSA.2025.0160908}
}