Simple and efficient methods for gait recognition using pose information

• Autor(a) principal: Vítor Cézar de Lima
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/38879; https://orcid.org/ 0000-0002-7552-1957
• topic: Gait recognition; Biometry; Computer vision; Computação – Teses; Visão por computador - Teses; Biometria – Teses; Reconhecimento de padrões (Computadores) – Teses

Gait is a biometry that differentiates individuals by their walking manner. Research on this topic has gained evidence since it is unobtrusive and available at distances, which is desirable in surveillance scenarios. Most of the previous works have focused on the 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 a pose estimation method, called PoseDist, to retrieve the coordinates of body parts and transform them into signals and movement histograms. These features are processed using a fusion of Subsequence Dynamic Time Warping and Euclidean distance to compare gait sequences from the probe with those in the gallery. This method is evaluated on all views of CASIA Dataset A and compared to existing ones, demonstrating its efficacy. However, as its algorithmic cost is high, it is only suitable for environments with few individuals; and this way, a new method called PoseFrame is employed for gait recognition, training a multilayer perception to classify poses from individual frames and aggregating its results by majority voting. PoseFrame is tested on CASIA Dataset A, having accuracy above other model-based works, including PoseDist; and on CASIA Dataset B, achieving state-of-the-art accuracy on same-view condition and having some of the best results on cross-view. Finally, an ablation study is also performed to find which body parts are the most important for gait recognition and according to the findings, the arms and feet are the most important locations.