An Improved Feature-Based Visual Slam Using Semantic Information

The traditional visual slam method for feature extraction is single and has poor robustness. This paper proposes an improved feature-based SLAM (Simultaneous localization and mapping) by adding weights to the features of objects matching the same semantic category and incorporating semantic information into loop closure detection. The basic idea is to use the deep neural network YOLOV5 to classify things, associate feature points with objects appearing in the bounding box, and thus assign the feature points to the semantic labels of these objects. In the feature matching process of the SLAM algorithm, the matching of feature points with the same semantic label will be weight, which increases the matching of similar features on the same category of objects, and at the same time intefrates semantic information into loop detection to improve the accuracy of loop closure detection. The test results show that the absolute trajectory error of the improved algorithm is more minor, the pose estimation accuracy is higher, and the tracking robustness can be effectively improved.