View-invariant gait person re-identification with spatial and temporal attention

Abstract

Person re-identification at a distance across multiple none overlapping cameras has been an active research area for years. In the past ten years, Short term Person Re-Id techniques have made great strides in terms of accuracy using only appearance features in limited environments. However, massive intraclass variations and inter-class confusion limit their ability to be used in practical applications. Moreover, appearance consistency can only be assumed in a short time span from one camera to the other. Since the holistic appearance will change drastically over days and weeks, the technique, as mentioned above, will be ineffective. Practical applications usually require a long-term solution in which the subject appearance and clothing might have changed after a significant period has elapsed. Facing these problems, soft biometric features such as Gait have been proposed in the past. Nevertheless, even Gait can vary with illness, ageing and changes in the emotional state, changes in walking surfaces, shoe type, clothes type, objects carried by the subject and even clutter in the scene. Therefore, Gait is considered a temporal cue that could provide biometric motion information. On the other hand, the shape of the human body could be viewed as a spatial signal which can produce valuable information. So, extracting discriminative features from both spatial and temporal domains would be very beneficial to this research. Therefore, this thesis focuses on finding the best and most robust method to tackle the gait human Re-identification problem and solve it for practical applications. In real-world surveillance scenarios, the human gait cycle is primarily abnormal. These abnormalities include but not limited to temporal and spatial characteristics changes such as walking speed, broken gait phase and most importantly, varied camera angles. Our work performed an extensive literature study on spatial and temporal gait feature extraction methods with a focus on deep learning. Next, we conducted a comparative study and proposed a spatial-temporal approach for gait feature extraction using the fusion of multiple modalities, including optical-flow, raw silhouettes and RGB images. This approach was tested on two of the most challenging publicly available datasets for gait recognition TUM-GAID and CASIA-B, with excellent results presented in chapter 3. Furthermore, a modern spatial-temporal attention mechanism was proposed and tested on CASIA-B and OULP datasets which learns salient features independent of the gait cycle and view variations. The spatial attention layer in the proposed method extracts the spatial feature maps using a two-layered architecture that are fused using late fusion. It can pay attention to the identity-related salient regions in silhouette sequences discriminatively using the spatial feature maps. The temporal attention layer consists of an LSTM that encodes the temporal motion for silhouette sequences. It uses the encoded output vectors in the temporal attention architecture to focus on the most critical timesteps in the gait cycle and discard the rest. Furthermore, we improved the performance of our method by mapping our extracted spatial-temporal gait features to a discriminative null space for use in our Siamese architecture for crossmatching. We also conducted an element removal experiment on each segment of our spatial-temporal attentional network to gain insight into each component’s contribution to the performance. Our method showed outstanding robustness against abnormal gait cycles as well as viewpoint variations on both benchmark datasets.