CBAM-SLAM: A Semantic SLAM Based on Attention Module in Dynamic Environment

Traditional visual SLAM systems are stable in static environments, but have low accuracy and unstable in dynamic condition, which affects positioning accuracy and mapping effects. To solve this problem, CBAM-SLAM is proposed in this paper which is a semantic SLAM system based on attention module in dynamic environment. First, we integrate the CBAM (Convolutional Block Attention Module) into the feature extraction network of Mask R-CNN to extract key features such as dynamic objects better and enhance the segmentation accuracy of dynamic objects. Subsequently, the masks generated by the improved Mask R-CNN network are input into the SLAM system as prior information. Next, the dynamic feature points are eliminated according to the prior information, and the static feature points are kept to tracking in VSLAM system. Finally, we conduct comparative experiment on the public dataset. The result shows that CBAM-SLAM effectively reduces the error of the real trajectory and the estimated trajectory, improves the accuracy and keep VSLAM system robust especially in high dynamic condition.