Kalman Filters and their application in motion detection and tracking

Abstract

During the last decade, breakthroughs in the fields of machine vision and reinforce-ment learning have been made possible thanks to the advancements in machine learning algorithms. ‘Machines’ now can recognize an object with human accuracy; this im-plies they could play video games by just having access to the pixels of a screen. In this work a combination of state-of-the-art technologies of deep learning are exploited to detect, track and remotely control a vehicle under realistic conditions. A linear estimator (Kalman filter) is proposed as an application-specific solution to a common problem that is the delay of sensory input – in the present case, the latency introduced by the machine vision system. The thesis presents the proposed system design and practical considerations along with detailed experimental results as a proof of concept. The advantages of the proposed solution are established through comparative evaluation with the results of other ‘naive’ solutions.