SEMANTIC MAPPING BASED ON VISUAL SLAM AND YOLOV5

Dawei Wang, Yuanwei Bi, Guohui Wang, and Yaowen Liu

Keywords

Visual SLAM, semantic map, RGBD image, online mapping

Abstract

Pose tracking and semantic mapping serve as the bedrock for enabling advanced environmental interaction by mobile robots in uncharted terrains. Tracking accuracy plays a pivotal role in determining the quality of the final map, while noisy image segmentation can introduce inaccuracies in the semantic details represented within the map. We synergistically leveraged the strengths of ManhattanSLAM and the YOLOv5 semantic segmentation network to achieve both tracking optimisation and globally consistent semantic mapping. Firstly, we enhanced the tracking accuracy through line feature optimisation and an optimised weight allocation strategy. This refinement ensured robust tracking performance. Next, we proposed a novel method for semantic mapping by combining surfels and supersets. This approach effectively addressed the issue of semantic inconsistency during the mapping. We performed extensive experiments across multiple datasets to validate the effectiveness of our algorithm. Specifically, evaluations on the TUM dataset revealed a 34% improvement in tracking accuracy compared to ManhattanSLAM, and demonstrated superior performance in producing high-quality semantic maps relative to other mapping methods.

Important Links:

Go Back