In injury-induced mortality among the elderly, falls represent one of the primary contributing factors. Thus, assessing fall-related risks holds critical significance for elderly health management. Existing fall risk assessments depend on instruments, clinical judgment, and environmental checks, but are often subjective and inconsistent. This study develops Bino-GaitRisk-Transformer, a novel binocular vision + Transformer architecture for fall risk prediction. By leveraging a confidence-weighted 3D reconstruction algorithm, it achieves millimeter-level joint localization accuracy (mean error \({<}10\) mm), reducing error by 40% compared to monocular systems. Unlike traditional models requiring handcrafted features, this approach directly uses raw gait data, with a hierarchical Transformer extracting spatial and temporal features via self-attention. The spatial module captures joint relationships, while the temporal module models gait dynamics, with cross-attention enabling deeper fusion. Experimental results demonstrated 92.6% classification accuracy and 0.918 F1-score, surpassing conventional approaches by \({>}20\) %. Ablation studies highlight the Transformer’s superior anomaly sensitivity (AUC = 0.926). This work innovatively applies Transformers to fall risk assessment, demonstrating an end-to-end framework integrating 3D skeletal data with deep learning to enhance elderly health monitoring.