Automatic multi-gait recognition using pedestrian's spatiotemporal features

This paper presents an automatic technique to detect and track the multiple pedestrians for their identifications in a video sequence. Contrarily to the most existing approaches, the proposed technique does not require human silhouette segmentation from the video to build the gait representation. Additionally, it also does not need to estimate the gait cycle to compute the gait-related features. The proposed technique comprises on four steps. In the first step, the pedestrian information is detected and tracked in the temporal direction. Second, we computed spatiotemporal features in the localized/tracked area to encode their walking patterns using dense trajectories. In the third step, the local features of pedestrian's walk are transformed into its compact and high-level representation using Fisher vector encoding scheme. Fourth, these high-level representations are fed to simple linear support vector machine for the identification. Since there is no publicly available multi-subject gait dataset and the recording of a new dataset is an expensive process which also demands a long time, we generated an augmented gait dataset where multiple subjects are available in a video sequence to cope with this limitation. We employed the single-subject CASIA-B gait dataset to generate the augmented multi-subject gait video sequences. The identification of multiple pedestrians in the constructed augmented gait sequences is a challenging task as multiple subjects are walking beside and crossing each other, hence producing several types of occlusions. The proposed gait recognition algorithm achieved a recognition rate of 86.3% on multi-subject gait dataset and 98.6% on the single-subject gait dataset.