In the implementation of autonomous driving systems, accurate acquisition of the vehicle’s location and orientation is crucial to providing a basis for path planning and obstacle avoidance. Although satellite navigation technology offers reliable positioning information, its signal is susceptible to interference in certain environments, such as areas with dense obstructions, which will affect the accuracy of vehicle localization and environmental mapping. To address the interference from dynamic points, we implement a method based on laser-vision multisensor fusion for identifying and extracting dynamic point clouds in environments where satellite signals are disrupted. We propose a dynamic point cloud extraction algorithm based on deep learning, utilizing semantic information to guide the network in extracting more precise information about the dynamic environment. We also construct a method for preliminary constraint of dynamic obstacles using global semantic information and design a multiframe point cloud processing approach with a sliding window mechanism, where residual data are accumulated and fed into the network as a composite model input. Experimental results demonstrate that our method significantly improves the positioning accuracy and map construction quality in dynamic environments, giving it considerable competitiveness compared with other advanced algorithms.