Post-integration based point-line feature visual SLAM in low-texture environments

To address the issues of weak robustness and low accuracy of traditional SLAM data processing algorithms in weak texture environments such as low light and low contrast, this paper first studies and improves the data feature extraction method, optimizing the AGAST-based feature extraction algorithm to adaptively adjust the extraction threshold according to the gradient size of different data features. Meanwhile, a fusion-based incremental loop closure detection method is proposed, which integrates the similarity scores of multi-dimensional data features based on the Borda counting strategy, thereby enhancing the accuracy of loop closure detection. The performance of loop closure detection was evaluated on public datasets (such as KITTI sequences 00, 05, and 06), achieving an average AP value of 92.03%. The overall system performance was evaluated on the EuRoC dataset, with the results showing a root mean square error range from 0.0061 to 0.0281 m, demonstrating the excellent accuracy and robustness of the proposed method in large-scale data processing.