In order to support fire emergency response in underground parking lots, this paper proposes a spatio-temporal adaptive risk-aware navigation method for fire scenes in parking lots. A computational framework integrating thermal imaging, smoke segmentation, lidar occupancy mapping, and graph search reweighting is constructed to characterize flame spread, visibility attenuation, and obstacle changes in real time. Experiments were carried out on a self-built dataset containing 12480 sets of thermal infrared-visible paired images and 3200 parking lot fire simulation scenes. The results show that the proposed method achieves 97.84% accuracy of fire target recognition, 95.31% recall rate of test set, and 34.6 ms inference delay under 320×320 input. In the path planning test, the task success rate reaches 98.6%, and the local replanning recovery delay is 0.73 s. In the two types of tasks, the global decision entropy of task A and task B is 0.912 and 0.934, respectively, and the average completion time is 23.4 s and 26.8 s, respectively, which shows that the proposed method has high stability and robustness in the parking environment with high occlusion, high thermal disturbance and dynamic blocking.
