TL;DR
This paper introduces a novel method for estimating motion fields from 3D LiDAR data, treating the entire scene as a collection of moving elements rather than tracking individual objects. This approach offers improved robustness and accuracy by mitigating issues inherent in traditional multi-target tracking, such as segmentation and data association errors.
What Problem Does This Paper Address?
Accurate motion perception is crucial for autonomous driving and collision avoidance systems. Traditional methods like Multi-Target Tracking (MTT) rely on segmenting and tracking individual objects (cars, pedestrians). However, MTT is susceptible to errors in object segmentation and data association, which can lead to inaccurate motion estimation. Visual motion field estimation (VMFE) offers an alternative by estimating motion for each point in the scene, but 2D VMFE lacks depth information, and 3D VMFE using cameras suffers from noise and environmental sensitivities. This paper addresses the need for a robust and accurate motion sensing method using 3D LiDAR, which provides direct 3D measurements and is less affected by environmental conditions, but requires new approaches to handle its unstructured point cloud data.
Methods
The proposed method, 3D LiDAR-based Motion Field Estimation (MFE), involves several key steps:
- Pre-processing: Raw 3D LiDAR point clouds are projected onto a polar grid. Ground points are removed as they are assumed static. Non-ground grids are then clustered into neighborhood systems.
- Grid Feature Extraction: For each grid, features like average height, height variance, and height difference are computed.
- Moving Grid Detection: Neighborhood systems are classified as "static" or "moving" based on their size. Only "moving" grids are considered for motion estimation.
- Data Association: A novel, fast data association algorithm is proposed. It uses a relaxation method for the Linear Integer Assignment Problem (LIAP) to reduce computational complexity from O(N³) to O(N²), enabling real-time performance. The cost function for association considers both position and feature differences between grids.
- Motion State Estimation: A Kalman filter is applied to estimate the motion state (position and velocity) of each moving grid. If a grid is associated with a previous scan, its filter is initialized with the previous state.
- Spatial Smoothing: To correct errors from data association and improve spatial coherence, an improved Least Mean Square (LMS)-based spatial smoother is introduced. This smoother, along with a mean smoothing step, refines the velocity estimates by considering the spatial correlation of motion among neighboring grids.
Datasets
The paper uses two datasets for validation:
- Wuhan Data Set: Collected by a Velodyne-64E LiDAR mounted on an intelligent driverless car at LuXiang Circle in Wuhan, China. This dataset features dense and sophisticated traffic with approximately 10-15 vehicles per scan.
- Karlsruhe Institute of Technology (KIT) Data Set: This dataset captures an ego-vehicle stopped at a cross-intersection with a car in front, leading to occlusion.
Key Insights and Ideas
- Scene-centric Motion Estimation: Shifting from tracking individual objects to estimating motion for every element in the scene (motion field) significantly reduces the impact of segmentation and data association errors, leading to greater robustness.
- Polar Grid Representation: Using polar grids for LiDAR data is advantageous as it naturally handles the decreasing point density with increasing range.
- Fast Data Association: The proposed relaxation method for LIAP is crucial for achieving real-time performance, a critical requirement for autonomous systems.
- Spatial Coherence for Robustness: Incorporating spatial smoothing based on the assumption that neighboring grids share similar motion properties effectively corrects localized errors and improves the overall quality of the motion field.
- LiDAR's Advantages: 3D LiDAR's direct 3D output and immunity to environmental conditions make it a superior sensor for robust motion sensing compared to cameras, despite the challenge of processing its unstructured data.
Key Results
- Accuracy: The proposed MFE method achieved an overall accuracy of 151.35 cm/s, which was better than the accuracy of the MTT algorithm and the method proposed by Moosmann and Stiller [5].
- Time Consistency: MFE significantly outperformed MTT in time consistency, with much smaller standard deviations in velocity differences between scans. This is attributed to MFE's avoidance of error-prone segmentation and its robustness to data association errors.
- Spatial Consistency: The LMS mean smoothing algorithm dramatically optimized the motion field, resulting in the smallest standard deviation of velocities within neighborhood systems, indicating excellent spatial consistency and removal of outliers.
- Real-time Performance: The average computation time for the proposed MFE algorithm was less than 100 ms (the time interval of data output), enabling it to run in real time. The time cost was found to be proportional to the number of moving grids.
Limitations and Future Work
- Limitations: The paper acknowledges that while the relaxation method for LIAP is locally optimal, it may not be globally optimal, potentially leading to some association errors. However, these are largely corrected by the spatial smoothing process. The accuracy of the method is still being evaluated, and further improvements are sought.
- Future Work: The authors plan to further improve the accuracy of motion state estimation and research methods for globally optimizing the motion field. They also aim to enhance the environmental sensing ability of intelligent vehicles and active collision avoidance systems.
Summary updated