Resumen
As consumer-grade depth sensors provide an efficient and low-cost way to obtain point cloud data, an increasing number of applications regarding the acquisition and processing of livestock point clouds have been proposed. Curve skeletons are abstract representations of 3D data, and they have great potential for the analysis and understanding of livestock point clouds. Articulated skeleton extraction has been extensively studied on 2D and 3D data. Nevertheless, robust and accurate skeleton extraction from point set sequences captured by consumer-grade depth cameras remains challenging since such data are often corrupted by substantial noise and outliers. Additionally, few approaches have been proposed to overcome this problem. In this paper, we present a novel curve skeleton extraction method for point clouds of four-legged animals. First, the 2D top view of the livestock was constructed using the concave hull algorithm. The livestock data were divided into the left and right sides along the bilateral symmetry plane of the livestock. Then, the corresponding 2D side views were constructed. Second, discrete skeleton evolution (DSE) was utilized to extract the skeletons from those 2D views. Finally, we divided the extracted skeletons into torso branches and leg branches. We translated each leg skeleton point to the border of the nearest banded point cluster and then moved it to the approximate centre of the leg. The torso skeleton points were calculated according to their positions on the side view and top view. Extensive experiments show that quality curve skeletons can be extracted from many livestock species. Additionally, we compared our method with representative skeleton extraction approaches, and the results show that our method performs better in avoiding topological errors caused by the shape characteristics of livestock. Furthermore, we demonstrated the effectiveness of our extracted skeleton in detecting frames containing pigs with correct postures from the point cloud stream.