Human gait recognition using topological information

Abstract

This paper shows an image/video application using topological invariants in human gait recognition. The 3D volume of a gait cycle is built stacking silhouettes extracted using a background substraction approach. Ideally, the border cell complex is obtained from the 3D volume with one connected component and one cavity. Then, it is necessary to apply a topological enrichment strategy in order to obtain a robust and discriminative representation for person recognition. Using a sliding cutter plane normal to some direction of view it is possible to divide the border cell complex in different parts. The incremental algorithm is used to compute the homology on each part. A vectorial representation is built ordering the number of connected components and tunnels obtained for each cut. In order to evaluate the robustness of this representation the silhouettes were diminished to a quarter of the original size. At the same time, this is considered a simulation of a human gait captured at long distance. Even, under these difficult conditions it was possible to get a 74% of correct classification rates on CASIA-B database.