An Appearance Invariant Gait Recognition Technique Using Dynamic Gait Features. (3rd May 2021)

• Record Type: Journal Article
• Authors: Masood, Hajra; Farooq, Humera
• Other Names: Mahmood Muhammad Tariq Academic Editor.
• Is Part Of: International journal of optics. Volume 2021(2021)
• Publication Date: 2021-05-03
• Subjects: Optics -- Periodicals; Optics; Periodicals; 535
• DOI: 10.1155/2021/5591728
• Languages: English
• ISSNs: 1687-9392

Abstract

Abstract : Gait recognition-based person identification is an emerging trend in visual surveillance due to its uniqueness and adaptability to low-resolution video. Existing gait feature extraction techniques such as gait silhouette and Gait Energy Image rely on the human body's shape. The shape of the human body varies according to the subject's clothing and carrying conditions. The clothing choice changes every day and results in higher intraclass variance and lower interclass variance. Thus, gait verification and gait recognition are required for person identification. Moreover, clothing choices are highly influenced by the subject's cultural background, and publicly available gait datasets lack the representation of South Asian Native clothing for gait recognition. We propose a Dynamic Gait Features extraction technique that preserves the spatiotemporal gait pattern with motion estimation. The Dynamic Gait Features under different Use Cases of clothing and carrying conditions are adaptable for gait verification and recognition. The Cross-Correlation score of Dynamic Gait Features resolves the problem of Gait verification. The standard deviation of Cross-Correlation Score lies in the range of 0.12 to 0.2 and reflects a strong correlation in Dynamic Gait Features of the same class. We achieved 98.5% accuracy on Support Vector Machine based gait recognition. Additionally, we develop a multiappearance-based gait dataset that captures the effects of South Asian Native Clothing (SACV-Gait dataset). We evaluated our work on CASIA-B, OUISIR-B, TUM-IITKGP, and SACV-Gait datasets and achieved an accuracy of 98%, 100%, 97.1%, and 98.8%, respectively.