Visual SLAM for Dynamic Environment Using Pre-Frame Semantic

To address the shortcomings of visual SLAM methods in dynamic environments, such as insufficient real-time performance and low robustness, this paper proposes a visual SLAM algorithm based on prior frame semantic information and DCP (Dilated Convolutional Pyramid) network layer. Firstly, to enhance the accuracy and speed of pixel semantic labeling in dynamic object regions, the target detection network employs the DCP network layer capable of fusing information from different-sized receptive fields, thus obtaining multi-scale features. Secondly, by integrating prior frame semantic information with inter-frame optical flow information using a memory module, the algorithm achieves pre-recognition of dynamic abnormal points in the current frame and fuses them with the environmental semantics identified by the target detection network, compensating for the omission of dynamic features due to semantic weakening. Experimental results show that the proposed algorithm effectively improves the removal of dynamic feature points, leading to more accurate pose estimation results. Furthermore, the semantic segmentation efficiency reaches over 60 frames per second, demonstrating better real-time performance.