Abstract
Robustness is an important criterion for evaluating semantic simultaneous localization and mapping (semantic SLAM). Limited to the computational cost and the complexity of the network, semantic information often has high uncertainty, this paper proposes a new method based on a continuous probability map, which integrates traditional feature points, semantic information, and historical data association. To be specific, an additional tread was added to the SLAM system. This thread utilizes an object detection network to generate high-level semantic information. On top of this semantic information, object detection was transformed into a novel probability map based on the Gaussian distribution assumption. This probability map also makes the optimizer focus on semantic-related objects while maintaining the background features. Further, we designed a continuous mechanism to update the probability map continuously. The mechanism was then applied to nonlinear optimization. The idea of probability map can be regarded as an individual module embedded into any other SLAM algorithm. The algorithm was tested on the TUM dataset. The results of the experiment show that, on average, our method improves localization accuracy by around 20% in indoor environments. Compared with many efficient SLAM systems including Dyna SLAM, the experimental results show that CP-SLAM achieves the effect of improving accuracy in the expected scenario.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call