In order to reduce the detection accuracy of pedestrians partially blocked by obstacles, this thesis proposes a semi-global stereo matching algorithm for partially blocked pedestrians based on multi-feature fusion. In this method, SLIC algorithm is used for super-pixel segmentation to improve pedestrian contour information, and multi-feature fusion tree path semi-global stereo matching algorithm is used to generate depth map. The pedestrian information, background information and obstacle information are separated by adaptive segmentation algorithm to obtain the area of interest. The area of interest was placed on the head and shoulders with obvious and stable pedestrian features to restrict the area of interest. Dimension reduction HOG is used to extract features and generate sample sets, and SVM classifier is trained to realize partially occlusioned pedestrian detection. Experiments show that compared with other pedestrian detection algorithms, the proposed algorithm has a higher pedestrian detection accuracy in the scene of partially occluded pedestrians, which proves the effectiveness of the proposed algorithm.