Gait Recognition from Occluded Sequences in Surveillance Sites

Abstract

Gait recognition is a challenging problem in real-world surveillance scenarios where the presence of occlusion causes only fractions of a gait cycle to get captured by the monitoring cameras. The few occlusion handling strategies in gait recognition proposed in recent years fail to perform reliably and robustly in the presence of an incomplete gait cycle. We improve the state-of-the-art by developing novel deep learning-based algorithms to identify the occluded frames in an input sequence and henceforth reconstruct these occluded frames. Specifically, we propose a two-stage pipeline consisting of occlusion detection and reconstruction frameworks, in which occlusion detection is carried out by employing a VGG-16 model, following which an LSTM-based network termed RGait-Net is employed to reconstruct the occluded frames in the sequence. The effectiveness of our method has been evaluated through the reconstruction Dice score as well as through the gait recognition accuracy obtained by computing the Gait Energy Image feature from the reconstructed sequence. Extensive evaluation using public data sets and comparative study with other methods verify the suitability of our approach for potential application in real-life scenarios. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.