Research on visual slam based on improved YOLOv5s in dynamic scenes

Abstract To address inaccuracies in camera position computations induced by dynamic objects in visual simultaneous localization and mapping (vSLAM), a method for dynamic feature point detection and rejection is given. This method merges an improved YOLOv5 lightweight network with the lucas-kanade (LK) optical flow. In the GhostNet network architecture, the typical convolutional layer is initially replaced by GhostConv separable convolution. At the end of the feature extraction network, a spatial attention mechanism based on CA is incorporated to safeguard crucial positional information and improve detection precision. This method efficiently eliminates dynamic elements, resulting in a point cloud map that is denser. Evaluations against the TUM dataset indicate that the proposed approach considerably surpasses previous SLAM methods regarding real-time efficiency and localization accuracy.