LiDAR-Inertial Odometry in Dynamic Driving Scenarios using Label Consistency Detection

TL;DR

A LiDAR-inertial odometry method that eliminates the influence of moving objects in dynamic driving scenarios is proposed, and achieves state-of-the-art performance on public datasets with extremlely low computational overhead.

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.

Figures (10)

• Figure 1: Illustration of (a) the exemplar point cloud map with dynamic points, where green points are ghost tracks of moving vehicles. (b) the static point cloud map, where the dynamic points have been detected and removed by our label consistency detection method.
• Figure 2: Illustration and examplar results of the proposed label consistency detection method. (a) Binarized labels are obtained by fast ground segmentation, where all points are divided into ground points (i.e., orange points in (a-2) and (a-3)) and non-ground points (i.e., white points in (a-2) and (a-3)). (b) Dynamic points are determined by comparing label consistency with nearest neighbors. The grid space where the dynamic point is located at the current time was not occupied in the past (i.e., green area in (b-1)). Thus the green points in (b-2) are detected as dynamic points because they are unable to find enough nearest neighbors. The occupied grid space does not exclusively contain static points (i.e., pink area in (b-1)), there may be some dynamic points adjacent to the ground. Thus the pink points in (b-2) are detected as dynamic points because their labels are inconsistent with surround ground points. The blue points and red points in (b-3) and (b-4) are ground and non-ground points in global map respectively.
• Figure 3: (a) Illustration of binarized label construction. The 3D points of current sweep are positioned in a range image, and a recursive 2D connected component method is utilized to identify ground points. (b) Visualization of segmented ground points from current input sweep. The orange points represent ground points and the white points represent non-ground points.
• Figure 4: (a) Illustration of the conventional 8-nearest neighbor search. All points within the 8-nearest neighbor voxels are compared with the distance to $\mathbf{p}^w$, and the 20 closest points are selected as the nearest neighbors (simply represented by 5 yellow points). (b) Illustration of the voxel-location-based nearest neighbor search. The voxel $V$ to which $\mathbf{p}^w$ belongs is directly located, and all points in $V$ (no more than 20) are considered as nearest neighbors (simply represented by 5 yellow points).
• Figure 5: Visualization of dynamic point determination results for fore-points. The areas obscured in white are the background regions, while those that remain visible are the foreground regions. The orange and white points represent the ground and non-ground points of current sweep, and the blue and red points correspond to the ground and non-ground points of global map. The green points are the identified dynamic points.