LiDAR-Inertial Odometry in Dynamic Driving Scenarios using Label Consistency Detection (2407.03590v2)

Abstract: In this paper, a LiDAR-inertial odometry (LIO) method that eliminates the influence of moving objects in dynamic driving scenarios is proposed. This method constructs binarized labels for 3D points of current sweep, and utilizes the label difference between each point and its surrounding points in map to identify moving objects. Firstly, the binarized labels, i.e., ground and non-ground are assigned to each 3D point in current sweep using ground segmentation. In actual driving scenarios, dynamic objects are always located on the ground. For most points scanned from moving objects, they cannot coincide with any existing structures in space. For a minority of moving objects' points that are close to the ground, their labels exhibit differences with surrounding ground points. Thus, the points on moving objects are identified due to lacking of nearest neighbors in map or inconsistency with the labels of surround ground points. The nearest neighbors from global map are localized by voxel-location-based nearest neighbor search and the consistency is evaluated by comparing the label consistency with nearest neighbors, without involving any massive computations. Finally, the points on moving objects are removed. The proposed method is embeded into a self-developed LIO system (i.e., Dynamic-LIO), evaluated with six public datasets, and tested in both dynamic and static environments. Experimental results demonstrate that our method can identify moving objects with extremlely low computational overhead (i.e., 1-9ms/sweep), and our Dynamic-LIO can achieve state-of-the-art pose estimation accuracy in both static and dynamic scenarios. We have released the source code of this work for the development of the community.

• The paper presents a label consistency-based detection method that rapidly identifies and removes dynamic points in LiDAR-Inertial Odometry systems.
• It integrates the method into a self-designed LIO framework using LeGO-LOAM for ground segmentation, ensuring real-time state estimation with reduced computational overhead.
• Empirical evaluations on benchmark datasets show competitive preservation and rejection rates while significantly enhancing pose estimation accuracy and processing speed.

An Expert Review of "A Fast Dynamic Point Detection Method for LiDAR-Inertial Odometry in Driving Scenarios"

The paper "A Fast Dynamic Point Detection Method for LiDAR-Inertial Odometry in Driving Scenarios," introduces a novel method for dynamic point detection and removal within the context of LiDAR-Inertial Odometry (LIO) systems. This approach addresses a crucial challenge in autonomous driving: the accurate and efficient detection of dynamic objects such as moving vehicles and pedestrians, which can impair localization and mapping accuracy.

Core Contributions

The authors present a method based on label consistency, which classifies LiDAR points depending on their consistency with neighboring point labels—distinguishing between ground and non-ground points. This technique allows the fast identification and removal of dynamic points, crucial for maintaining the reliability of LIO systems in dynamic environments.

1. Label Consistency-Based Detection: The primary contribution is a label consistency-based method to efficiently detect dynamic points. Points without a sufficient number of nearest neighbors, or those with label discrepancies compared to their neighbors, are classified as dynamic.
2. Integration with LIO: The method is embedded within a self-designed LIO system, which focuses on enhancing real-time state estimation accuracy while minimizing computational overhead. The use of ground segmentation from LeGO-LOAM ensures resource-efficient point classification.
3. Source Code Availability: By providing the codebase open-source, the authors promote further development and testing within the research community.

Evaluation and Results

Empirical validation is performed on well-known datasets, including Semantic-KITTI, ULHK-CA, and UrbanNav. The proposed method exhibits commendable preservation and rejection rates, akin to existing state-of-the-art methods, but with remarkable improvements in computational efficiency.

• Preservation (PR) and Rejection Rates (RR): The method achieves competitive PR and RR with existing methods like Removert and Erasor, demonstrating robust static point preservation alongside effective dynamic point removal.
• Pose Estimation Accuracy: The integration of the method into an LIO system notably enhances pose estimation accuracy, surpassing current state-of-the-art systems such as LIO-SAM, RF-LIO, and ID-LIO, especially in highly dynamic driving scenarios.
• Computational Efficiency: The algorithm processes dynamic point detection and removal within 1 to 9 milliseconds, significantly outpacing previous methods like Dynamic Filter and RH-Map. This efficiency is pivotal for maintaining the real-time operation of LIO systems in autonomous vehicles.

Implications and Future Directions

The method's low computational overhead is particularly relevant for the deployment in resource-constrained platforms typical of autonomous systems. The approach's reduced time consumption while preserving competitive performance indicates its potential for widespread adoption in real-time applications.

From a theoretical standpoint, this work contributes to the growing body of literature exploring the integration of sensor fusion with enhanced dynamic environment modeling. It sparks opportunities for refined methods that balance computational efficiency with robustness against false positives in dynamic cluttered environments.

While the results are promising, future research could explore:

• Exploring adaptive models that dynamically balance computational resources according to scene complexity.
• Extending the framework to cater to multi-sensory data fusion scenarios, incorporating visual or radar data for a more holistic environmental understanding.
• Developing adaptive thresholds for dynamic point classification to improve robustness across diverse operational environments.

In conclusion, the paper sets forth a compelling method for dynamic point detection in LiDAR-inertial systems, achieving notable gains in computational efficiency and accuracy which underline its practical applicability in the field of autonomous navigation.