Person Re-Identification (Pre-id) in Blur, Low Light, Low Resolution Surveillance Videos: Object Detection and Characterization of Sequences
Manisha Talware, Sanjay Mahadev Koli · 2021 International Conference on Smart Generation Computing, Communication and Networking (SMART GENCON) · 2021
Person re-identification (Pre-id) aims at matching an individual across different cameras, time and place. Pre-id has moved from an image-based system to a video-based system and datasets are moving from closed datasets to real-world surveillance videos. Collection of datasets and annotating them is one of the challenges. Deep Learning (DL) outperforms conventional feature extraction and matching methods used for Pre-id. Cloud platforms such as Google Colab ease out the burden of huge resource requirements of DL algorithms. Problem with pedestrian detection, classical object detection is at the root of Pre-id. This paper focuses on Pre-id methods trends in datasets and annotation methods and providing information at-a-glance. It further discusses pedestrian detection using Deep CNN architecture for differently characterized sequences and tries to correlate the results with model and sequence. The results outperform the detection task. It achieves detection of all, except maximally occluded persons with an average probability of more than 90% in varied situations. This paper contributes to establish a foundation for the next task of Pre-id. It compiles datasets and annotation information at a glance along with the information on methods and trends in Pre-id.