Suspect Tracking through Person Re-Identifcation

Abstract

Person Re-Id tackles the query of locating a specifc an individual in many images or videos, sometimes taken with various cameras in various locations. This thesis gives a complete framework for re identifying a person in a camera network in order to automate the surveillance system. Attention neural models are used in the current state-of-the-art solutions. Combination of different loss functions on top of a temporal attention-based neural network model and apply bag of trick over it to outperform current state of art person re-identifcation results on PRID2011 dataset. Our combined loss function is combination of Center Loss (CL) and Online Soft Mining Loss (OSM) summation with Attention Loss (AL). As the need for surveillance and camera networks has expanded in past few years, the use of video to re-identify people has become a critical task and has attracted a lot of interest. A typical video-based person Re-Id system consists of an image-level feature extractor (such as CNN), a temporal modelling method for aggregating temporal information, and a loss function. Although several temporal modelling methods were described, direct comparisons are challenging since the used feature extractor and the used loss function have signifcant effect on the fnal result. For suspect tracking through person Re-Id over PRID2011 and self generated custom dataset, the proposed model is Temporal Attention plus bag of tricks strategy