A novel deep neural design and efficient Pipeline architecture for Person Re-Identification in high resolution Video

The primary objective of person re-identification (Re-ID) is to retrieve a person of interest across different nonintersecting cameras for managing in distributed surveillance systems. This has added to its increasing popularity on account of its widespread use, applications and research significance. In this study, we have proposed a novel pipelined deep learning architecture which acts as a robust feature extractor and also helps in reducing down the search space by generating feature embeddings followed by executing a distance metric measurement for finding the similar neighbourhood embeddings and subsequently sorting the cluster centroids of the matching embeddings for finding a set of the nearest match before passing down to a siamese network for similarity checking in the reidentification process. Our experiments reported to achieve a performance accuracy of 85~90% with a model size of 288 MB executing at 30 fps in real-time.