LWDNet: Low-Altitude Wire Detection Network based on Forward-Looking Images of Aircraft

Power lines and high-voltage cables pose significant safety threats to aircraft operating in low-altitude airspace. Ensuring accurate and comprehensive perception of these aerial obstacles has become critical for enhancing aviation safety and improving the ability of airworthiness. In this paper, a Low-altitude Wire Detection Network (LWDNet) is proposed for aerial power line segmentation in low-altitude environments. Compared with the general object detection task, the challenges of aerial wire detection arise from the following aspects: long-span distribution, high aspect ratio, inconsistent visual patterns, and low-resolution signatures. Furthermore, flight scenarios introduce additional complexities from aircraft attitude variations, viewing angles, occlusions, and dynamic background environments, resulting in low intra-class similarity. To address the above challenges, this paper designs a Dynamic Snake Bottleneck (DSB) module that enhances feature attention on elongated targets through adaptive curvature modeling. Moreover, we propose a Frequency Fusion Residual Decoding Strategy (FFRS) that reconciles multi-spectral features to mitigate appearance inconsistency while improving algorithmic robustness. The proposed LWDNet is rigorously evaluated on both the public TTPLA dataset and our self-constructed WDOA dataset. Experimental results demonstrate that LWDNet achieves superior wire segmentation performance in low-altitude scenarios, with high Intersection over Union (IoU) scores (83.64%) and low missed detection rates (6.02%), significantly outperforming existing stateof-the-art methods.