An Improved Deep-Layer Architecture for Real-Time End-to-End Person Recognition System

• Published in: Computers & electrical engineering 2021-12, Vol.96, p.107550, Article 107550
• Main Authors: Jayavarthini, C, Malathy, C
• Format: Article
• Language: English
• Subjects: Automation; Cameras; Conventional neural network; Deep-learning; End-to-end; Faster-RCNN; Feature map; High-order relationships; Human detection; Neighbourhood differences; patch features; Person recognition; Person search; Real time; Recognition; Spatial differences; Surveillance; video surveillance
• ISSN: 0045-7906; 1879-0755
• DOI: 10.1016/j.compeleceng.2021.107550

•Finding matches for a person in a real-time surveillance video is challenging.•Challenges include changes in background, illumination, pose, occlusion, views etc.•Assigning unique identifier to same person is known as person re-identification.•Deep learning with customized layers captures robust features of person's image.•Similarity score is computed using the spatial differences across summary of patches.•Threshold value is fixed to remove irrelevant matches. Surveillance of human activities in real time has drawn tremendous attention in the field of research. As manual monitoring of surveillance videos is expensive and prone to error, automation of surveillance is preferred. Person recognition is one of the fundamental problems related to automation of surveillance. It is defined as the system that generates correspondence between two images captured by different cameras at different times. Matching of probe image with the people in the surveillance video is really challenging due to variations in background, costume of people, pose, camera views, lighting, etc. A deep-learning-based end–end person recognition system is proposed to suit the real-world environment. This paper discusses the architecture of proposed system with the issues encountered during the implementation. Experiments were conducted based on different situations to illustrate the results of the proposed system with suitable evaluation metric. CUHK03 dataset was used for experiment. Real-time data were collected and tested to prove the robustness of the proposed system. [Display omitted]