Matching Gait Image Sequences in the Frequency Domain for Tracking People at a Distance

Abstract

Abstract. This paper describes a new method to track people walking by matching their gait image sequences in the frequency domain. When a person walks at a distance from a camera, that person often appears and disappears due to being occluded by other people and/or objects, or by going out of the field of view. Therefore, it is important to track the person by taking correspondence of the image sequences between before and after the disappearance. In the case of tracking, the computational time is more crucial factor than that in the case of identification. We create a three-dimensional volume by piling up an image sequence of human walking. After using Fourier analysis to extract the frequency characteristics of the volume, our method computes the similarity of two volumes. We propose a method to compute their correlation of the amplitude of the principal frequencies to improve the cost of comparison. Finally, we experimentally test our method and validate that the amplitude of principal frequencies and spatial information are important to discriminate gait image sequences.

Multi-view large population gait dataset and its performance evaluation for cross-view gait recognition

“…Thus, evaluating such problems is a future work. Though we evaluated only GEI-based methods in order to compare gait recognition methodologies, there are many gait features that can be used for input images, such as DeepGait [41], CGI (chrono-gait image) [42], and FDF (frequency-domain feature) [43]. Thus, evaluating such gait features is also a future work.…”

Dynamic FingerPrint based on human motion and posture

“…Sagawa et al [19] performed matching of gait image sequence in the frequency domain. A volume was created by piling up the image sequences of walk.…”

Multiple views gait recognition using View Transformation Model based on optimized Gait Energy Image

“…The upper half of silhouettes are aligned along horizontal centroid [2]. Then, each bounding box of silhouette is W × H. 1D Fourier Transform is carried out at each position of (x, y) on S n,t (x, y) across all frames (along the time sequence) [7] [8]. Then FG is extracted by computing 1D Fourier transform along a time axis based on gait periodicity [7] [8].…”