Multimodal Feature Association-based Stereo Visual SLAM Method

Much work has been done to improve visual SLAM systems by integrating point, line, and plane features into the bundle adjustment model, however little attention has been paid to explore the associations between these spatial features that could be used to achieve better performance. This study proposes a multimodal feature association-based stereo SLAM method. Firstly, a method to extract point and line features from stereo images and estimate plane features through the line features is proposed. Then, the corresponding mathematic models for representing the association relations between different features are given. After that, the association relations are integrated into the back-end optimization model to improve system robustness and accuracy by adjusting reprojection errors with confidence weights. Finally, comparison tests are performed to evaluate the performance of the proposed method with the mainstream stereo SLAM systems using the EuRoC and KITTI datasets. The test results show that our approach not only generates semantic maps with higher fidelity but also provides a better positioning capability, with the positioning accuracy of the algorithm on the EuRoC and KITTI datasets improved by an average of 14.78% and 20.09%, respectively.