Autonomous Driving: Mapping and Behavior Planning for Crosswalks

Abstract

As autonomous driving integrates with every day traffic, early adopters are initially skeptical and designers are overly cautious. With safety as the top priority, current systems are sometimes too slow at executing maneuvers. Scenarios such as switching into a crowded lane or waiting for a left turn can result in the autonomous system to wait much longer than a human driver would. This behavior can be frustrating for passengers and confusing for other drivers around. Acceptable driving style also depends on other context like location and culture. A driver may be more forceful in a densely populated city compared to a calmer driver from the suburbs. This thesis explores the unsignalized pedestrian crosswalk scenario and methods that balance safety, assertiveness, caution, and obstruction of traffic flow when interacting with pedestrians. A configurable driving policy for the Autonomoose system is introduced with results. The work adopts the lanelet mapping format and introduces a method of mapping and representing the crosswalk regulation. The main contribution of the work is a tunable algorithmic approach for progressing through unsignalized crosswalks that exemplifies both conservative and assertive driving behavior. The algorithm described in this work is one of possibly infinitely many methods for handling unsignalized crosswalks. Reinforcement learning based solutions and other hand crafted algorithms can benefit from using the work proposed as a point of comparison. General concepts proposed in the algorithm may inspire more robust algorithms in future development.