Observation Time Difference: an Online Dynamic Objects Removal Method for Ground Vehicles

TL;DR

This paper tackles the challenge of removing dynamic objects from LiDAR maps in urban environments by introducing an online dynamic removal method that does not require a prior map. It classifies dynamic objects into suddenly appear and suddenly disappear using the observation-time difference with the ground, and employs downward retrieval, upward retrieval, and static restoration within a voxel-based map framework. The approach achieves strong real-time performance (≈$23.8$ ms per frame) and competitive or superior robustness on SemanticKITTI and challenging datasets, with a reported speedup of over 60% compared to state-of-the-art online/offline methods. Limitations include reduced performance on uneven terrain and potential failures with aerial objects due to reliance on ground-ground relationships.

Abstract

In the process of urban environment mapping, the sequential accumulations of dynamic objects will leave a large number of traces in the map. These traces will usually have bad influences on the localization accuracy and navigation performance of the robot. Therefore, dynamic objects removal plays an important role for creating clean map. However, conventional dynamic objects removal methods usually run offline. That is, the map is reprocessed after it is constructed, which undoubtedly increases additional time costs. To tackle the problem, this paper proposes a novel method for online dynamic objects removal for ground vehicles. According to the observation time difference between the object and the ground where it is located, dynamic objects are classified into two types: suddenly appear and suddenly disappear. For these two kinds of dynamic objects, we propose downward retrieval and upward retrieval methods to eliminate them respectively. We validate our method on SemanticKITTI dataset and author-collected dataset with highly dynamic objects. Compared with other state-of-the-art methods, our method is more efficient and robust, and reduces the running time per frame by more than 60$\%$ on average.

Figures (5)

• Figure 1: The results of dynamic objects removal. The upper image is original map. The lower image is static map constructed by our method. The middle image is partially enlarged image, and the purple area is traces left by dynamic objects.
• Figure 2: Overview of our proposed method. Our proposed method includes three modules: ground segmentation, map management and dynamic removal. The point cloud $\mathbf{P}$ is divided into ground point set $\mathbf{G}$ and non-ground point set $\mathbf{U}$ in ground segmentation module, and then is sent to map management module and dynamic removal module together with the input pose $\mathbf{T}^W$. After downward retrieval, upward retrieval and static restoration, dynamic objects can be recognized.
• Figure 3: Dynamic objects classification. Only ground was observed at the begining in (a). But after a period of time, both ground and non-ground object were observed. The object is a suddenly appear dynamic object. On the contrary, both ground and non-ground object were observed at the beginging in (b). But after a period of time, only ground was observed. The object is a suddenly disappear dynamic object.
• Figure 4: Downward retrieval. The image on the left represents the $i$-th LiDAR frame, while the image on the right represents the $j$-th LiDAR frame ($j > i$). The points observed by LiDAR are represented by red dots. The object was at position E at the $i$-th LiDAR frame, and it moved to position C at the $j$-th LiDAR frame. We first observed the non-ground voxel (C, 4) in the $j$-th LiDAR frame, but we had already observed the ground voxel (C, 2) below it in the $i$-th frame. Therefore, we compare the minimum LiDAR frame index between these two voxels to determine whether (C, 4) is dynamic or not.
• Figure 5: The dynamic removal results in challenging scenarios, the partial enlarged image is shown below. The purple points in the picture is traces left by dynamic objects, and the area with serious error removal is in orange boxes.