Socially compliant robust navigation in crowded pedestrian environment

Abstract

This thesis describes a robot navigation system that relies only on onboard sensors without a high-definition map to achieve real-time perception and planning in a crowded pedestrian environment. Integration of real-time perception and planning with interactive decision-making is the leading research focus of this project. The perception module detects the drivable area in real-time using a semantic segmentation network, estimates the state of surrounding pedestrians using an object detection and tracking network, and predicts the future state of pedestrians by a pedestrian prediction network. The planning module integrates the perception information and calculates the robot's trajectory. Two methods based on optimization and sampling are applied, and both planning performances are compared. We also develop a pedestrian simulator to verify the functionality of the navigation system and then deploy the navigation system on a real robot. Experiments show that the robot can easily handle common pedestrian navigation scenarios and even some more complex scenarios.