With the growing aging population, the demand for reliable fall detection systems is increasing. Owing to safety concerns, real-world fall data—especially for tripping-induced falls—are often limited in existing studies. To address this limitation, we developed a large-scale fall-protection system and conducted controlled experiments specifically targeting tripping falls, ensuring subject safety throughout. Subjects wore two inertial measurement units mounted on the torso—one on the chest and the other on the abdomen—and performed both tripping falls and fall-like behaviors. These wearable inertial sensors continuously captured torso motion signals, providing time-series acceleration and angular velocity data for subsequent analysis. On the basis of the collected time-series data, three long short-term memory-based models were developed: two single-sensor models and one dual-sensor model. A leave-one-subject-out cross-validation approach was applied, and a low sigmoid decision threshold together with a consecutive-window decision rule was adopted to reduce missed detections. Experimental results demonstrated high classification performance across all models, with the dual-sensor model achieving the best accuracy, precision, recall, and F1-score. These findings confirm the effectiveness of the proposed method in distinguishing tripping falls from similar daily activities.