View-invariant gait recognition exploiting spatio-temporal information and a dissimilarity metric

Abstract

In gait recognition, when subjects do not follow a known walking trajectory, the comparison against a database may be rendered impossible. Some proposed solutions rely on learning and mapping the appearance of silhouettes along various views, with some limitations caused for instance by appearance changes (e.g. coats or bags). The present paper discusses this problem and proposes a novel solution for automatic viewing angle identification, using minimal information computed from the walking person silhouettes, while being robust against appearance changes. The proposed method is more efficient and provides improved results when compared to the available alternatives. Moreover, unlike most state-of-the- art methods, it does not require a training stage. The paper also discusses the use of a dissimilarity metric for the recognition stage. Dissimilarity metrics have shown interesting results in several recognition systems. This paper also attests the strength of a dissimilarity-based approach for gait recognition.