In unknown or hazardous environments (such as emergency rescue and disaster relief), traditional navigation methods struggle to achieve specific target navigation due to the inability to obtain prior maps and location information. This paper proposes a target-driven autonomous navigation method for mobile robots based on dueling double deep Q network (D3QN). The cross-modal fusion module of this method dynamically weights and integrates features from different modalities, effectively consolidating observational data while fully capturing environmental information. This significantly enhances the capability to perceive the environment.Building on this, a general target-driven navigation approach is designed, where YOLOv5 is used to recognize specific targets (such as flames or smoke) and obtain their locations. The identified target locations are used to replace predefined waypoints in deep reinforcement learning-based navigation, enabling autonomous navigation to specific targets. Simulation results show that the proposed method has significant advantages in the navigation success rate and other indicators. In simple, complex, and dynamic scenarios, the success rates increased by 9%, 27%, and 38%, respectively. Moreover, the model trained in a simple simulation environment can be directly deployed in more complex simulation environments and real-world scenarios, exhibiting strong generalization capability.