关键词: 低空经济, 无人机, 视觉定位, 状态空间模块, 自监督训练, Airsim, 特征匹配

Abstract:

Drones in the Urban Air Mobility （UAM） environment are faced with the problems of unstable satellite navigation and positioning and signal interference， and existing additional navigation systems are faced with the challenges of requirement for high airborne computing power. This study proposes a lightweight UAV visual navigation and positioning system. By developing an image feature extraction framework that fuses state space modules， the prediction accuracy is significantly improved， and a triplet self-supervised training method based on Gaussian pyramid is implemented to enhance the robustness of the algorithm. By introducing a feature matching strategy based on sliding windows and similarity matrices， the feature matching process is optimized and the inference speed is significantly improved. Experiments on the Airsim simulation platform and real UAM flight scenarios show that this algorithm can provide accurate additional navigation and positioning data in complex environments. Multiple sets of ablation experiments and performance tests are conducted to verify the advanced nature and real-timeliness of the algorithm， as well as its capability to effectively reduce the computing power requirements. The results show that this system can not only improve the accuracy of UAV navigation， but also provide a feasible visual solution for positioning and navigation in urban air traffic environments.

Key words: low-altitude economy, UAV, visual positioning, state space module, self-supervised training, Airsim, feature matching