uPLAM: Robust Panoptic Localization and Mapping Leveraging Perception Uncertainties
CoRR(2024)
摘要
The availability of a reliable map and a robust localization system is
critical for the operation of an autonomous vehicle. In a modern system, both
mapping and localization solutions generally employ convolutional neural
network (CNN) –based perception. Hence, any algorithm should consider
potential errors in perception for safe and robust functioning. In this work,
we present uncertainty-aware panoptic Localization and Mapping (uPLAM), which
employs perception uncertainty as a bridge to fuse the perception information
with classical localization and mapping approaches. We introduce an
uncertainty-based map aggregation technique to create a long-term panoptic
bird's eye view map and provide an associated mapping uncertainty. Our map
consists of surface semantics and landmarks with unique IDs. Moreover, we
present panoptic uncertainty-aware particle filter-based localization. To this
end, we propose an uncertainty-based particle importance weight calculation for
the adaptive incorporation of perception information into localization. We also
present a new dataset for evaluating long-term panoptic mapping and map-based
localization. Extensive evaluations showcase that our proposed uncertainty
incorporation leads to better mapping with reliable uncertainty estimates and
accurate localization. We make our dataset and code available at:
更多查看译文
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要