Gait Recognition Using Pose Estimation and Signal Processing

Abstract

Gait is a biometry that differentiates individuals by the way they walk. Research on this topic has gained evidence since it is unobtrusive and can be collected at distance, which is desirable in surveillance scenarios. Most of the previous works have focused on human silhouette as representation. However, they suffer from many factors such as movement on scene, clothing and carrying conditions. To avoid such problems, this work employs pose estimation to retrieve the coordinates of body parts, which are transformed into signals and movement histograms to be used as feature descriptors. While the former descriptors are used with the Subsequence Dynamic Time Warping that compares signals from probe and gallery, the Euclidean distance is used on the latter to find the person on gallery that is closest to probe. Finally, the outputs of both are fused. This work was evaluated on all views of CASIA Dataset A and compared to existing ones, demonstrating its efficacy.