Aircraft skin pit damage detection algorithm based on multiscale surfaces

To address the problems of strong noise interference, long detection times, uneven fuselage surfaces, lack of visual information in two-dimensional images, and difficulty in automatic detection, an automatic detection algorithm for aircraft skin pit damage based on a multiscale surface model was designed. First, an automatic acquisition platform system was constructed using an unmanned vehicle, a lifting pole, and a depth camera. The point-cloud data of the aircraft skin was obtained using this acquisition platform system. The point-cloud data were then preprocessed using the radius filter algorithm, voxel grid filter algorithm, and moving least squares algorithm. Then, the preprocessed point-cloud data were divided into multiscale regions and split into multiple local skin mesh regions to obtain multiple local grid area data. For each local grid region data, the surface models of each local grid region and regional spatial adjacency were obtained by constructing and optimizing the estimation of the local quadric surface based on the random sampling consensus algorithm. The spatial adjacency, surface model, and its index together form the region tree. The local surface models at different scales were aggregated by storing information and normal vector angles in the region tree to identify the damaged and nondamaged regions. Finally, surface features, such as curvature and normal vector, were used to cluster the pit points in the damaged area, and the pit point-cloud data were aggregated to obtain the final pit damage results. The proposed algorithm was compared with existing traditional algorithms, such as the point-cloud block method and the point feature region growth method based on normal vector and curvature. Experimental results showed that the accuracy, recall, and F-value of the point-cloud block method were 4.00%, 20.00%, and 6.67%, respectively, with an average detection time of 20 s. For the point feature region growth method based on normal vector and curvature, the accuracy, recall, and F-value were 30.77%, 26.67%, and 28.79%, respectively, with an average detection time of 25 s. The accuracy, recall, F-value, and average detection time were significantly improved, with mean values of 92.86%, 86.67%, 89.92%, and 6 s, respectively. Additionally, the detection results of the three algorithms on the aircraft skin engine, fuselage, and wing were compared, and the influence of curvature in different regions on the algorithm was analyzed. The detection results of the proposed algorithm were significantly better than those of existing traditional algorithms, such as the point-cloud block method and the point feature area growth method based on normal vector and curvature. The proposed algorithm achieved the goal of automatically detecting pit damage in aircraft skin scenes.