Abstract: In response to the decline in detection performance of existing pedestrian detection algorithms in crowded scenes, due to occlusions, scale variations, and complex environmental interferences, an improved pedestrian detection algorithm based on CenterNet is proposed. By integrating the attention mechanism and the asymmetric pyramid non- local block module into the backbone network, the feature extraction capability and the ability to capture contextual information were enhanced, thus improving the detection effectiveness for occluded targets. A dual- branch neck network was employed to fuse features of different scales, enhancing the detection accuracy for small- scale targets. Experimental results demonstrate that the proposed algorithm outperforms the traditional CenterNet algorithm and current mainstream detection algorithms on the CityPersons and CrowdHuman datasets, achieving accurate detection of occluded and small- scale pedestrians.