Data Association Network

Abstract

In this thesis we present a novel approach of developing correspondence cost matrix to improve the data association phase by fusing data from multiple sources. Our cost matrix is composed of both position and visual cues and provides better performance than existing online state of the art trackers. The proposed algorithm provides more than realtime performance and can be deployed in computationally restrictive environments. Furthermore vectorized implementations of the proposed algorithm can allow for large scale real-time tracking which makes it for industrial usage. To the best of my knowledge no real time online tracker provides better performance trade off than the proposed algorithm. The employed technique seamlessly blends deep learning, a more advanced approach with Hungarian Algorithm, a rather primitive cost assignment algorithm, in order to improve the performance of tracking in a tracking by detection framework. The algorithm has been benchmarked on the MOT 17 Challenge, which is a standard benchmarking tool for all tracking algorithms and has alot of challenging sequences.