Deep Convolutional Multi-Relational Graph Attention Network for Autonomous Drone Navigation in Urban Wind Zones

Drones are essential for civil engineering operations like logistics and data collecting. Current autonomous drone studies mainly concerns itself with safe path planning in static scenarios; however one of the major challenges of over urban environment is often dynamic factors such as wind, building layouts, and signal coverage that can significantly impact direction stability. For autonomous navigation across dynamic wind zones, this study proposed a new Shape-aware Mesh Dual Lightweight Deep Convolutional Multi-Relational Graph Attention Network with Growth Optimizer (SAMDLDCMRGANet-GO) that leverages local data from RGB cameras and GPS. The method incorporates a Dual Aggregation Transformer for feature extraction, Shape-aware Mesh Normal Filtering for preprocessing, and the LDCMRGAN et for efficient path planning. The proposed approach demonstrates its efficacy in navigating complicated surroundings while retaining high precision in obstacle avoidance and target acquisition, achieving an impressive 99.14% recall, 99.15% accuracy, an outstanding EVA of 0.99, and a low RMSE of 0.41.