A vision-based water leakage detection method using a quadruped robot for concrete structures

Concrete structures are indispensable to modern transportation and infrastructure, yet water leakage remains a critical threat to their safety and durability. Seepage reduces load-bearing capacity, compromises operational safety, increases maintenance costs, and causes environmental pollution. Existing leakage-detection methods rely mainly on manual inspection, which is inefficient, costly, and heavily dependent on human factors. To overcome these limitations, a quadruped-robot-based visual leakage-detection approach is introduced. First, the noise characteristics of images acquired under dynamic locomotion are analyzed, and a preprocessing pipeline synergistically integrates Wiener filtering, wavelet decomposition, and particle-swarm optimization is developed for denoising. Subsequently, a lightweight coupled architecture is designed by cascading the YOLOv8 backbone with HorNet, BoTNet, and MSFN modules, on this basis, a semantic-segmentation algorithm for water leakage identification is built. Finally, the efficacy of the proposed method is demonstrated through a leakage dataset collected on-site by the quadruped robot. The results reveal that the proposed method exhibits strong detection performance, with the mean pixel accuracy (mPA) reaching 93.71 %. The proposed method accurately segments leakage regions from the background, offering a valuable benchmark for high-precision, autonomous leakage detection in complex and confined environments.