To address the problems of insufficient scene perception ability and decoupling of the decision-making process from environmental dynamics when a mobile robot performs end-to-end navigation tasks in an unknown environment, we propose a navigation algorithm based on the Transformer architecture for the coupled representation of the target and the scene. First, RGB images from the visual sensor are fused with the physical state information of the target and then embedded and encoded. With FastViT (Fast Vision Transformer) as the core (serving as the main component of the perception module), local and global feature extraction is performed on the input data to learn and generate a scene feature representation that integrates target semantics. Second, a multi-modal input framework is constructed by incorporating LiDAR data to enhance the robot's ability to understand complex scenes, and the SAC (Soft Actor-Critic) deep reinforcement learning algorithm is used for action decision-making. Finally, a reward function for the deep reinforcement learning algorithm is designed based on the safety risk capsule. By dynamically quantifying collision risk in path planning, the safety of the navigation process is improved. Simulation experiments are carried out on the GAZEBO platform. The results show that, compared with mainstream algorithms in the same field, the proposed algorithm achieves a 17.36% increase in average navigation success rate and effectively avoids obstacles. This algorithm provides a new reference for the safe navigation of mobile robots in unknown environments.