To address the problems of a large parameter size and insufficient detection accuracy in complex scenarios for multi-scenario elderly fall detection, a multi-scenario elderly fall detection based on an improved YOLOv12 was proposed. First, grouped shuffle convolution (GSConv) was embedded into the backbone network of YOLOv12, while the Triplet attention mechanism was introduced and the neck network was replaced with the lightweight Slim Neck structure. Second, the improved YOLOv12 model was trained and validated on a self-built fall detection dataset. Finally, detection experiments were conducted on the test set to evaluate the detection accuracy and robustness of the improved model. Experimental results show that, compared with the original YOLOv12 model, the parameter size of the improved model is reduced from 256 MB to 234 MB, a decrease of 85.9%, and the FLOPs are reduced from 63 G to 54 G, a decrease of 1429%. Meanwhile, mAP@50 increases from 83.8% to 87.6%, mAP@50-95 increases from 51.3% to 55.4%, and Recall increases from 76.5% to 81.3%. The improved YOLOv12 model achieves a favorable balance between lightweight architecture and detection accuracy, providing an effective solution for multi-scenario elderly fall detection.