Evidential Semantic Mapping in Off-road Environments with Uncertainty-aware Bayesian Kernel Inference

TL;DR

Comprehensive experiments across various off-road datasets demonstrate that the proposed evidential semantic mapping framework enhances accuracy and robustness, consistently outperforming existing methods in scenes with high perceptual uncertainties.

Abstract

Robotic mapping with Bayesian Kernel Inference (BKI) has shown promise in creating semantic maps by effectively leveraging local spatial information. However, existing semantic mapping methods face challenges in constructing reliable maps in unstructured outdoor scenarios due to unreliable semantic predictions. To address this issue, we propose an evidential semantic mapping, which can enhance reliability in perceptually challenging off-road environments. We integrate Evidential Deep Learning into the semantic segmentation network to obtain the uncertainty estimate of semantic prediction. Subsequently, this semantic uncertainty is incorporated into an uncertainty-aware BKI, tailored to prioritize more confident semantic predictions when accumulating semantic information. By adaptively handling semantic uncertainties, the proposed framework constructs robust representations of the surroundings even in previously unseen environments. Comprehensive experiments across various off-road datasets demonstrate that our framework enhances accuracy and robustness, consistently outperforming existing methods in scenes with high perceptual uncertainties.

Figures (5)

• Figure 1: In perceptually challenging unstructured outdoor environments, considering semantic uncertainty from the segmentation network is beneficial to produce more accurate semantic maps by prioritizing confident predictions. By incorporating the semantic uncertainty of the network, our uncertainty-aware semantic BKI mapping framework can produce reliable and accurate 3D semantic maps in off-road scenes.
• Figure 2: Overview pipeline of our uncertainty-aware semantic BKI framework. With an evidential segmentation network trained by EDL, input data is processed to derive continuous semantic probability and uncertainty. These 3D semantic points are then integrated into the semantic map through Bayesian updates using the uncertainty-aware BKI, resulting in a dependable semantic map and variance map in uncertain off-road environments.
• Figure 3: Qualitative results of 3D semantic mapping methods. Compared to others, our method generates reliable and accurate maps that preserve semantic details while excluding noisy predictions. In RELLIS-3D, only our method reconstructs grass consistently (First row), and dirt roads and puddles in detail (Second row). In our OffRoad dataset, our method accurately reconstructs the boundaries of unpaved roads, grass, and vegetation, compared to others.
• Figure 4: Zero-shot semantic mapping results on our OffRoad dataset using a semantic segmentation network pre-trained on RUGD. Our method robustly constructs semantic maps despite prediction uncertainties in unseen environments, whereas other methods struggle to produce clear maps.
• Figure 5: Visualization of variance maps. Through uncertainty-aware BKI, our method generates more informative variance maps capable of reliably quantifying semantic map uncertainty.