Machine Learning Meets Localization

Abstract

This thesis project was conducted in cooperation with Zenseact for the purpose of creating a solution for determining the lane in which an autonomous vehicle is driving. Solving this is part of the larger problem of localization and state estimation of autonomous vehicles and is referred to as Lane-Level Localization (LLL). The problem is connected to the area of Early Time Series Classification, which is the field of applying supervised learning and time series classification techniques for classifying time series accurately with as few observations as possible. The problem of LLL may be solved by applying what is known as a multi-hypothesis technique. This is a technique that estimates some state by tracking several different possibilities (hypotheses) for the state and using some model for inferring the most likely scenario. It is found that using an architecture that allows for the possibility of rejecting output depending on the certitude with which a classification can be made can be adapted to solving the problem of LLL in autonomous vehicles. In the current scenario, the model produced an accuracy of 99,5% while only rejecting to classify in 1% of the sequences.