Based on the thermal infrared characteristics, the infrared stereo vision pedestrian perception method can effectively detect and measure pedestrians in road scenes at night and hazy environments, with the aim of improving driving safety. Owing to less texture details in infrared images, the traditional dense binocular stereo matching algorithm performs poorly. To solve this problem, the region of interest (ROI) is extracted according to the brightness and edge features of the targets in the infrared image. Then, the image feature points are extracted and matched in the ROI to calculate the original sparse depth map. Finally, according to the small depth difference in the surface of the targets, the semi-dense depth map was estimated by combining the ROI and the original depth map. We designed an experimental system to verify the effectiveness of the proposed method. The experimental results showed that the relative error of the depth perception of pedestrians was better than 1.5% at 15 m and 3% at 30 m in the field of view of approximately 120°.