An algorithm for detecting partially occluded pedestrians based on multi-feature fusion and tree-structured semi-global stereo matching was proposed to address the issue of reduced detection accuracy in pedestrian detection caused by partial obstruction. The simple linear iterative clustering (SLIC) algorithm was employed for superpixel segmentation to enhance the contour information of pedestrians, and the tree-structured multi-feature fusion semi-global stereo matching algorithm was used to generate depth maps. Pedestrian, background, and obstacle information were separated using an adaptive segmentation algorithm to obtain the region of interest. The region of interest was positioned around the head and shoulders of the pedestrian, where features were distinct and stable, to impose constraints. Feature extraction was conducted using dimension-reduced histogram of gradient (HOG), and a sample set was generated for training an support vector machines (SVM) classifier, ultimately achieving the detection of partially occluded pedestrians. The experiment shows that compared with other pedestrian detection algorithms, the proposed algorithm has a higher accuracy in pedestrian detection in partially occluded scenes, proving the effectiveness of the proposed algorithm.