Unsupervised Vehicle Re-Identifi cation

Abstract

Vehicle re-identi cation (Re-ID) is one of the primary components of an au- tomated visual surveillance system. It aims to automatically identify/search vehicles in a multi-camera network usually having non-overlapping eld-of- views. Majority of approaches dealing with the re-ID problem tackle it in a supervised manner which have certain limitations that pose challenges of generalization e.g., large amount of annotated data is required for training and is often limited to the dynamic growth of the data. Unsupervised learn- ing techniques can potentially cope with such issues by drawing inference directly from the unlabeled input data and have been e ectively employed in the context of person re-ID. Vehicle re-ID can also be performed using license plate recognition as its the unique ID of a vehicle. But there ex- ist some issues e.g., low resolution, environmental factors, side viewpoints, poor/improper illumination, partially broken, and fainted and/or occluded etc. To this end, this thesis presents an unsupervised approach to solve the vehicle re-Id problem by training a base (deep) network architecture with a self-paced progressive unsupervised learning technique. Such a cascaded formation has been successfully employed in the context of person re-ID but, to our knowledge, has not been applied to solve the vehicle re-ID problem. Moreover, the presented approach incorporates the contextual information into the proposed progressive framework that signi cantly improves the con- vergence of the learned algorithm. The analysis of the developed algorithm has been extensively analyzed over two large available benchmark datasets VeRi and VehicleID for vehicle re-ID with image-to-image and cross-camera searches and achieved better performance in most of the standard evaluation metrics when compared with the existing state-of-the-art approaches.