A Robust Multi-Camera Deep Person Re Identification Framework Using Spatiotemporal Context Modelling

Abstract

A Robust Multi-Camera Deep Person Re-Identification Framework Using Spatiotemporal Context Modelling In this thesis, research work is based on a robust automatic re-identification of a person from multiple non-overlapping cameras under variable and dynamic environmental conditions for an accurate re-identification and retrieval of targeted person identities. Person Re-Identification (ReID) aims at identifying query person of interest (POI) assigned with a unique identity label across multiple non-overlapping cameras. The query POI can be either an image or a video sequence. Person ReID has gained quite increasing attention among various research and developer communities in recent years. Several research challenges including occlusion, variable viewpoint, misalignment, unrestrained poses, background clutter, etc. are the major challenges in developing robust, lightweight, end-to-end trainable person ReID models. To address these issues, an attention mechanism that comprises local part/region aggregated feature representation learning is presented in this research by incorporating long-range local and global context modeling. The part-aware local attention blocks are aggregated into the widely used modified pre-trained ResNet50 CNN architecture as a backbone employing two attention blocks i.e. Spatio-Temporal Attention Module (STAM) and Channel Attention Module (CAM) thus improving both local and global feature representation learning. The spatial Attention block of STAM can learn contextual dependencies between different human body parts regions like head, upper body, lower body, and shoes from a single frame. On the other hand, the temporal attention modality is capable to learn temporal contextual dependencies of the same person’s body parts across all video frames. Lastly, the channel-based attention modality i.e. CAM can model semantic connections between the channels of feature maps. These STAM and CAM blocks are combined sequentially from a unified attention network named Spatio-Temporal Channel Attention Network (STCANet) that will be able to learn both short-range and long-range global feature maps respectively. Extensive experiments are carried out to study the effectiveness of STCANet on three images and two video-based benchmark datasets i.e. Market- x 1501, DukeMTMC-ReID, MSMT17, DukeMTC-VideoReID, and MARS. K reciprocal re-ranking of gallery set is also applied in which the proposed network showed significant improvement over these datasets in comparison to the state-of-the-art by achieving (mAP/Rank-1) score of (95.5/94.5), (90.7/92.3) and, (74.4/84.5) on Market-1501, DukeMTMC-ReID, and MSMT17 dataset respectively. In addition, the proposed modified STCANet also showed significant performance improvement in comparison to state-of-the-art methods by achieving (mAP/Rank-1) score of (96.6/97.1), (85.3.7/89.1) on DukeMTMC-VideoReID and MARS dataset respectively Lastly, to study the generalizability of STCANet on unseen test instances, cross-validation on external cohorts is also applied that showed the robustness of the proposed model. The proposed STCANet is lightweight, end-to-end trainable, and can be easily deployed to the real world for practical applications