Abstract: Pedestrian detection algorithm is a basic algorithm to support the strong and safe operation of autonomous driving system. Due to the limitations of realistic application scenarios such as occlusion and scale variation, existing algorithms often consider constructing more complex feature extraction structures, and lack of consideration of reasoning delay and detection recall. To solve this problem, this paper proposes a real-time robust pedestrian detection algorithm for complex scenes such as occlusion. A low-parameter multi-scale feature fusion block was designed to capture multi-scale pedestrian feature information with low computational overhead. An attention block based on strip receptive field was designed to highlight the local correlation of pedestrians in the feature map, so as to improve the reasoning accuracy in the case of occlusion. Experimental results in crowded and small-scale scenes show that compared with the network before optimization, the detection accuracy of the proposed algorithm is improved, and the network complexity and deduction delay are greatly reduced, which realizes low delay and high robust pedestrian detection in complex scenes such as occlusion.