Robust gait based human identification on incomplete and multi-view sequences

Abstract

Gait based person identification is an important research area in the field of video surveillance. The major challenges faced by gait recognition systems in real-life scenarios include view variance, occlusion and resultant unavailability of a complete sequence containing a gait cycle. In this work, we propose a novel robust gait recognition framework capable of handling these challenges. We show how Gait-Energy-Images (GEIs) can be accurately constructed from largely incomplete input silhouette sequences. This provides an immediate advantage over current literature that assumes availability of complete sequences. We then highlight the shortcoming of most of the current view-invariant models that perform sub-optimal transformation of probe and gallery sequences captured in different views for comparison. We propose a model which jointly estimates and learns the optimal transformation for comparison of probe and gallery GEIs. Through extensive experiments, we show that our proposed framework is able to outperform most state-of-the-art methods on multiple benchmarks.