HiMo: High-Speed Objects Motion Compensation in Point Clouds

TL;DR

The paper addresses non-ego motion distortions in LiDAR point clouds caused by fast-moving objects, which are especially problematic in highway and multi-LiDAR setups. It introduces HiMo, a pipeline that repurposes self-supervised scene flow estimation to undistort dynamic regions, coupled with SeFlow++ to improve scene flow under high-speed conditions. Two novel evaluation metrics—compensation accuracy at the point level and object shape similarity—assess non-ego motion compensation, with extensive experiments on Scania highway data, Argoverse 2, and ZOD demonstrating improvements in geometric consistency and downstream perception tasks. The work provides open-source code and velocity-aware ground-truth strategies, highlighting substantial benefits for segmentation and 3D detection when processing HiMo-corrected data.

Abstract

LiDAR point cloud is essential for autonomous vehicles, but motion distortions from dynamic objects degrade the data quality. While previous work has considered distortions caused by ego motion, distortions caused by other moving objects remain largely overlooked, leading to errors in object shape and position. This distortion is particularly pronounced in high-speed environments such as highways and in multi-LiDAR configurations, a common setup for heavy vehicles. To address this challenge, we introduce HiMo, a pipeline that repurposes scene flow estimation for non-ego motion compensation, correcting the representation of dynamic objects in point clouds. During the development of HiMo, we observed that existing self-supervised scene flow estimators often produce degenerate or inconsistent estimates under high-speed distortion. We further propose SeFlow++, a real-time scene flow estimator that achieves state-of-the-art performance on both scene flow and motion compensation. Since well-established motion distortion metrics are absent in the literature, we introduce two evaluation metrics: compensation accuracy at a point level and shape similarity of objects. We validate HiMo through extensive experiments on Argoverse 2, ZOD, and a newly collected real-world dataset featuring highway driving and multi-LiDAR-equipped heavy vehicles. Our findings show that HiMo improves the geometric consistency and visual fidelity of dynamic objects in LiDAR point clouds, benefiting downstream tasks such as semantic segmentation and 3D detection. See https://kin-zhang.github.io/HiMo for more details.

Figures (13)

• Figure 1: Multi-LiDARs are equipped in our heavy vehicles to avoid self-occlusion. (a) shows an example placement with 6 LiDARs. The point colors in (b-c) correspond to the LiDAR from which the points are captured. (b) illustrates the distortion of static structure due to fast-moving ego vehicle. Raw shows the raw data, w. egc shows the ego-motion compensation results. (c) demonstrates distortion caused by motion of other objects, which depends on the velocity of the said objects. In such case, ego-motion compensation alone (w. ego-motion comp.) is insufficient. In comparison, our HiMu pipeline (w. HiMo motion comp.) successfully undistorts the point clouds completely, resulting in an accurate representation of the objects.
• Figure 2: Distribution of object speeds in different datasets. Only objects with speeds exceeding 2m/s are included in the plot. The orange line shows the median speed and the blue dashed lines indicate the $\pm10\%$ spread.
• Figure 3: Top-view example of LiDAR sweeps showing how distortions are created for vehicles with (a) a single LiDAR and (b) two LiDARs. The small concentric circles of yellow and green are LiDARs. Both cases cause a displacement distance for the high-speed object. The light and dark cars show the vehicle's positions at two timestamps (the beginning and the end of the LidDAR sweep). (a) One complete single LiDAR scan sweep (small yellow dots are the first and last scan points). This case is only observed when moving objects are at the scan boundaries. (b) Two LiDAR scans separated in orientation by 180 degrees. This case is always observed for fast-moving objects. An animation illustrating both cases is included in the supplementary video.
• Figure 4: Examples of LiDAR distortion in various datasets after ego-motion compensation Argoverse2Waymoalibeigi2023zenseact. Ground points are removed in visualizations for a clearer view. Each column shows an example from a different dataset. Within each column, the top image shows the full scene, while the three images below show the zoomed-in front, top-down, and side views, respectively. (a) and (b) showcase the scenario described in \ref{['fig:reason_1']}, where an object is captured right at the beginning and at the end of the scan in a single LiDAR setup. (c) and (d) showcase the multi-LiDAR distortion scenario described in \ref{['fig:reason_2']}. The different colors in these two subfigures represent data from different LiDARs.
• Figure 5: Schematic of the HiMo pipeline. Given a sequence of consecutive point cloud frames, a scene flow estimator is employed to calculate the flow of each point. This flow, together with the known LiDAR scan interval and point time difference $\Delta T(\mathbf{p})$, allows us to compute the 3D distortion correction vector. Finally, the undistorted point cloud is computed by combining the correction vector with the raw point cloud.