S$^2$MAT: Simultaneous and Self-Reinforced Mapping and Tracking in Dynamic Urban Scenariosorcing Framework for Simultaneous Mapping and Tracking in Unbounded Urban Environments

TL;DR

S$^2$MAT is a novel framework that enables robots to simultaneously map static environments and track dynamic objects in large-scale, unstructured urban scenarios, improving navigation and perception in real-time.

Contribution

It introduces an integrated approach combining dynamic object tracking with static mapping, leveraging their reciprocal relationship for enhanced performance in highly dynamic environments.

Findings

Achieves state-of-the-art dynamic detection and tracking accuracy.

Demonstrates robust and scalable navigation over 7 km in urban scenarios.

Provides high-quality static maps despite dynamic obstacles.

Abstract

Despite the increasing prevalence of robots in daily life, their navigation capabilities are still limited to environments with prior knowledge, such as a global map. To fully unlock the potential of robots, it is crucial to enable them to navigate in large-scale unknown and changing unstructured scenarios. This requires the robot to construct an accurate static map in real-time as it explores, while filtering out moving objects to ensure mapping accuracy and, if possible, achieving high-quality pedestrian tracking and collision avoidance. While existing methods can achieve individual goals of spatial mapping or dynamic object detection and tracking, there has been limited research on effectively integrating these two tasks, which are actually coupled and reciprocal. In this work, we propose a solution called S$^2$MAT (Simultaneous and Self-Reinforced Mapping and Tracking) that integrates a front-end dynamic object detection and tracking module with a back-end static mapping module. S$^2$MAT leverages the close and reciprocal interplay between these two modules to efficiently and effectively solve the open problem of simultaneous tracking and mapping in highly dynamic scenarios. We conducted extensive experiments using widely-used datasets and simulations, providing both qualitative and quantitative results to demonstrate S$^2$MAT's state-of-the-art performance in dynamic object detection, tracking, and high-quality static structure mapping. Additionally, we performed long-range robotic navigation in real-world urban scenarios spanning over 7 km, which included challenging obstacles like pedestrians and other traffic agents. The successful navigation provides a comprehensive test of S$^2$MAT's robustness, scalability, efficiency, quality, and its ability to benefit autonomous robots in wild scenarios without pre-built maps.