Escalator fall detection algorithms in subway stations are a crucial means of preventing passenger accidents. However, in a real scene, issues such as the loss of small objects due to scale changes and interference caused by complex backgrounds can lead to false positives and false negatives. This paper proposes a passenger escalator fall detection algorithm based on SCGD-Yolo11m-pose network. First, RFD module was introduced during the downsampling stage to improve the robustness of feature extraction. Second, in C2PSA, deformable attention was used and named C2DA. This enables the model to enhance its ability to perceive various falling poses in complex backgrounds. Finally, in neck network, Gold-Yolo structure replaces the PANet network to strengthen the recombination ability of multi-scale features and improves the accuracy of the model in complex background. Additionally, only four keypoints defined by COCO, both shoulders and both hips, are retained, which improves computational efficiency. Experimental results on our self-built subway escalator fall dataset show that the improved model is improved by 2.3% on AP50 and 2.5% on AP50:95. This validates the effectiveness and practicality of the proposed algorithm in ensuring the safety monitoring of subway passengers.