An Improved Monocular SLAM System by Combining Geometric Features With Semantic Probability

Monocular SLAM has gained widespread popularity due to its numerous advantages. However, this approach is susceptible to appearance changes and often experiences problems such as sensitivity of dynamic objects and scale drift. Using semantics as an advanced visual feature is highly effective since it remains unaffected by appearance and viewpoint changes. This makes it easier to establish a more robust data association. In this paper, we represent the semantic information of features as the semantic probability that they belong to each category, thus reducing the adverse effects of semantic segmentation noise and increasing the ability of semantic representation. Furthermore, we propose a method that combines the semantic probability with geometric features to enhance the accuracy and robustness of the SLAM system. Specifically, the odometry module chooses geometric-semantic feature point descriptors, which help achieve higher accuracy in feature matching and eliminate the influence of dynamic features. It also provides an initial estimation of the real scale. The back-end optimization module utilizes semantic probability to establish medium-term constraints, effectively mitigating the cumulative drift of attitude and scale. The loop detection module constructs global and local descriptors using semantic information and incorporates the DBOW2 algorithm based on geometric features, enabling accurate and robust loop detection. The experimental results demonstrate that the proposed algorithm achieves high-performance pose estimation on the test sequences compared to other advanced monocular SLAM methods. Compared with DBOW2 based on appearance, the AUC (area under the curve) of our loop detection method is increased by 28.1% while remaining lightweight.