Abstract
TOSS (Tracking and moving Object Segmentation for Static scene mapping) is a novel real-time framework that enables autonomous robots to simultaneously track moving obstacles for safe navigation and build clean, static maps. By introducing a hierarchical association cost matrix and a spatio-temporal voting scheme, the system achieves a significant speedup while maintaining high mapping accuracy even in challenging, unstructured environments.
Introduction
In autonomous robotics, Simultaneous Localization and Mapping (SLAM) and navigation are interdependent. Safe navigation requires detecting and tracking moving objects, while reliable mapping requires filtering those same objects to avoid "ghosting" artifacts. Historically, these tasks have been treated separately, resulting in either accurate but slow offline map cleaning or fast but low-quality real-time navigation maps. TOSS aims to provide a unified, computationally efficient framework that handles both tasks in real-time.
Methodology
The TOSS framework consists of three primary stages designed for efficiency and robustness:
- Segmentation (TRAVEL): The system uses a spherical projection-based algorithm to separate raw LiDAR point clouds into "ground" and "above-ground" instances, represented as 3D bounding boxes.
- 3D Multi-Object Tracking (3DMOT): To track these boxes, TOSS utilizes a Kalman filter with a Hierarchical Association Cost Matrix. Unlike standard tracking that compares every detection to every track ($O(N^2)$ complexity), TOSS only calculates costs for the $k$-nearest neighbors, reducing computational complexity to $O(N)$.
- Dynamic-Static Voting (DS-Voting): To handle tracking errors like flickering labels due to occlusions, DS-Voting analyzes the spatio-temporal history of an object. If an object appears in the same spatial area across non-continuous frames, it is "voted" back into the static category, significantly reducing false positives in the final map.
Experimental Evaluation
The authors evaluated TOSS using the SemanticKITTI benchmark and a Real-world Unstructured Dataset collected on the KAIST campus using Unitree quadruped robots. This real-world data is particularly challenging due to steep hills, stairs, and dense vegetation, where the "bumpy" motion of a walking robot often breaks traditional SLAM algorithms.
Key Results and Insights
TOSS demonstrates superior performance in both speed and map quality:
- Efficiency: In the "Duckpond" environment, TOSS achieved a runtime of 0.029s per frame, a nearly 10x speedup over the Auto-MOS baseline, making it viable for low-power embedded systems.
- Map Preservation: On SemanticKITTI sequence 00, TOSS achieved a 99.87% Preservation Rate (PR), compared to just 34.57% for OctoMap.
- Robustness: The system successfully removed dynamic objects and recovered static structures in complex scenarios where traditional ground-plane assumptions fail.
Conclusion
TOSS provides a robust, real-time solution for tracking and mapping in dynamic environments. By leveraging spatio-temporal memory through DS-Voting, it effectively cleans maps of "ghost" artifacts while remaining resilient to the erratic movements of legged robots. Future work will focus on making the initial instance segmentation more robust to over- and under-segmentation errors.
Summary updated