Visual SLAM Based on Lightweight Dynamic Object Detection

The traditional visual synchronous localization and mapping algorithm (visual SLAM) assumes that the external environment is static. The system positioning accuracy drops when it is applied in the dynamic environment. In this paper, CBAM (Convolutional Block Attention Module) and ShuffleNetV2 are fused to improve the network feature extraction ability, and the fused network is used to reconstruct the backbone network of YOLOv5s, reducing the amount of model parameters and improving the algorithm reasoning speed. The improved object detection network and optical flow method are combined to eliminate the dynamic feature points extracted from the front end of the SLAM. And the remaining static feature points are used for inter frame matching to solve the camera pose. The experimental results on TUM dataset show that this method has improved the positioning accuracy of the system in indoor dynamic environment, and has achieved some improvement in real-time and accuracy compared with other classical dynamic visul SLAM algorithms.