Pose-based boundary energy image for gait recognition from silhouette contours

Gait is biometric that refers to the walking style of an individual, and medical studies claim that every individual has a distinctive gait pattern. In this work, we consider a gait recognition scenario in which at least a half cycle of gait is captured from the fronto-parallel view for each person and propose a new feature termed the Pose-based Boundary Energy Image that captures the dynamics of gait by extracting features from the silhouette contour information corresponding to several fractional parts of a gait cycle. Traditionally, binary silhouettes extracted from the input RGB frames of a gait sequence were used to construct the gait features for recognition. In contrast, silhouette contour information for feature extraction, as used in this work, is advantageous in the sense that it makes the extracted gait features more discriminative due to eliminating the redundant pixel-level structural information. We next reduce the dimension of the extracted Pose-based BEI feature using PCA, and finally carry out LDA-based classification using the reduced feature set. Evaluation of the proposed approach has been done using the popular CASIA-B and TUMGAID data sets and quite satisfactory results are obtained. A comparative study with existing techniques also shows that our approach outperforms the existing methods both in terms of average accuracy and classification time.