With the growing urban traffic flow, group pedestrian crossings have become increasingly common, demanding higher perception capabilities from Autonomous vehicles in complex environments. However, mutual occlusion among pedestrians significantly degrades the performance of traditional detection methods in crowded scenes, posing risks to autonomous vehicle safety. To solve this problem, this paper proposes a Multi-modal Pedestrian Detection under Occluded scenes (MPGDO) model based on improved machine vision. Designed within a two-stage detection framework, MPGDO utilizes a dual-channel feature encoding structure to extract both head and body features simultaneously. It further integrates head features with an Extendable Non-Maximum Suppression (ENMS) strategy, which dynamically adjusts the IoU threshold to optimize bounding box selection and enhance detection stability. Evaluated on the CityPersons dataset and a self-collected Beijing Zao Yuan Road dataset, MPGDO demonstrates effective detection across all occlusion levels (“heavy” to “no occlusion”). In particular, it reduces the false positive rate to 9.59% under moderate occlusion and achieves a 6.37% accuracy improvement over the YOLOv10 baseline, while maintaining strong robustness even under severe occlusion. These results validate its effectiveness in detecting pedestrians in complex, occluded traffic scenarios.