TL;DR
This paper introduces an efficient online method for removing dynamic objects from LiDAR maps generated by ground vehicles. By analyzing the observation time difference between objects and the ground, it effectively distinguishes and removes "suddenly appearing" and "suddenly disappearing" dynamic objects, leading to cleaner maps with significantly reduced processing time.
What Problem Does This Paper Address?
Creating accurate and reliable maps is fundamental for autonomous robots, especially for localization and navigation. However, urban environments are filled with dynamic objects like cars and pedestrians. When mapping with LiDAR, these dynamic objects leave behind "traces" or ghost-like artifacts in the point cloud map. These artifacts can severely degrade the performance of subsequent tasks, such as localization (determining the robot's position) and path planning, by creating false obstacles or confusing the robot's understanding of its surroundings.
Existing methods for removing these dynamic objects often operate "offline," meaning they process the map after it has been fully constructed. This adds significant computational overhead and delays, making them unsuitable for real-time applications where maps are continuously updated. The paper aims to bridge this gap by proposing an "online" dynamic object removal method that works concurrently with the mapping process, providing clean maps in real-time.
Methods
The core innovation of this paper lies in its novel approach to identifying dynamic objects based on the observation time difference between an object and the ground it's on. The method operates in three main modules:
- Ground Segmentation: This module accurately separates ground points from non-ground points. It uses a two-step process: initial filtering followed by PCA refinement to ensure a clean ground point set, which is crucial for the subsequent steps.
- Map Management: A voxel-based map is maintained, divided into three submaps: ground, non-ground, and dynamic. Each voxel stores not only points but also the LiDAR frame indices when it was observed. This allows for tracking the observation history of each voxel.
- Dynamic Removal: This is the heart of the method. It classifies dynamic objects into two categories:
- Suddenly Appear: These objects are detected when a non-ground voxel is observed, but the ground voxel directly below it was observed earlier. The method uses "downward retrieval" to check if the non-ground voxel's first observation time is significantly later than the ground voxel's first observation time (beyond a threshold
Tret). - Suddenly Disappear: These objects are detected when a ground voxel is observed, but the non-ground voxel that was previously above it is no longer observed. The method uses "upward retrieval" to check if the ground voxel's last observation time is significantly later than the non-ground voxel's last observation time (beyond
Tret). - Static Restoration: A final step corrects for potential misclassifications of static objects as dynamic, especially in scenarios with noisy LiDAR data or complex ground surfaces. It compares the total observation times of potentially dynamic voxels with their corresponding ground voxels.
Datasets
The proposed method was evaluated on:
- SemanticKITTI dataset: A widely used benchmark for semantic scene understanding in LiDAR point clouds, which includes dynamic object annotations.
- Author-collected dataset: A custom dataset designed to include highly dynamic scenarios.
Key Insights and Ideas
- Observation Time Difference as a Dynamic Object Signature: The paper's central insight is that the temporal relationship between an object's appearance/disappearance and its supporting ground surface is a strong indicator of dynamism. This is a more robust cue than relying solely on motion or comparing against a static prior map.
- Online Processing without a Prior Map: By focusing on the temporal differences within the current frame and recent history, the method avoids the need for a pre-built static map, enabling true online operation.
- Classification into "Appear" and "Disappear" for Targeted Removal: Separating dynamic objects into these two categories allows for specialized retrieval strategies (downward vs. upward), making the removal process more precise.
- Ground Contact is Key: The assumption that dynamic objects are typically in contact with the ground is leveraged effectively for both classification and retrieval.
Key Results
- Efficiency: The method significantly reduces processing time, achieving an average reduction of over 60% per frame compared to other state-of-the-art methods. The dynamic removal module itself takes only a few milliseconds.
- Robustness: The method demonstrates stable performance across different segments of the SemanticKITTI dataset, outperforming other online methods.
- Effectiveness: Qualitative results show successful removal of dynamic object traces, preserving important static details. In challenging scenarios with high pedestrian density, the method achieved more significant dynamic object removal than existing approaches.
- Comparable Performance: While not always achieving the absolute highest F1 scores (especially against offline methods like ERASOR2 that use instance segmentation), the proposed method offers a strong balance of performance and real-time capability.
Limitations and Future Work
- Reliance on Ground Contact: The paper acknowledges that the method's performance might degrade in uneven terrain scenarios where the assumption of ground contact is less reliable.
- Inability to Detect Aerial Objects: The method is designed for ground vehicles and thus cannot detect aerial dynamic objects.
- Future Work: The authors suggest exploring improvements for uneven terrain and aerial object detection.
Summary updated