UAV real-time visual navigation and obstacle perception based on dual-attention mechanisms and multimodal fusion
摘要
With the widespread application of unmanned aerial vehicle (UAV), the real-time perception and dynamic obstacle avoidance capabilities of visual navigation systems in complex environments have emerged as critical bottlenecks limiting autonomous operations. In dense obstacle scenarios, traditional visual navigation algorithms often fail due to insufficient feature extraction or temporal modeling deviations, proving inadequate for precise avoidance requirements in dynamic environments. This paper proposes DualAtt-DepthViT, a deep visual algorithm for UAV navigation that addresses challenges in obstacle perception and real-time control by integrating dual-attention mechanisms with multimodal weighted fusion. First, the spatial attention module dynamically focuses on obstacle edges and critical regions in depth images, significantly enhancing refined perception capability of local features. Simultaneously, the hierarchical weighting of historical motion states through hierarchical spatiotemporal attention LSTM ensures temporal consistency between short-term emergency avoidance and long-term path planning. Second, we introduce a multimodal adaptive fusion strategy that optimizes weight allocation for image, position, and attitude data through end-to-end learning, effectively mitigating information redundancy and conflicts inherent in conventional uniform fusion approaches. Furthermore, leveraging the global spatial modeling capacity of vision transformer (ViT) combined with the dynamic temporal prediction advantages of LSTM, our framework establishes an end-to-end mapping from obstacle semantic understanding to motion control. Experiments in the Flightmare simulation scenario show that the success rate of obstacle avoidance of this algorithm is improved by