A Mobile Cloud Collaboration Fall Detection System Based on Ensemble Learning
Tong Wu, Yang Gu, Yiqiang Chen, Jiwei Wang, Siyu Zhang · 2020 · Proceedings of the 22nd International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS) · doi:10.1145/3373625.3417010
Summary
This paper addresses fall detection for older adults, which the WHO identifies as the second leading cause of accidental injury death worldwide. Roughly 28-35% of people over 65 experience falls each year, with the rate increasing to 32-42% for those over 70. The authors propose a two-part solution: a novel machine learning algorithm called Fall-detection Ensemble Decision Tree (FEDT) and a three-stage mobile cloud collaboration system architecture. The system uses a smartwatch (Samsung Gear S3) equipped with accelerometers as the mobile device. In the first stage, a lightweight threshold method on the watch filters out normal activities of daily living (ADLs) by calculating the root mean square of tri-axial accelerometer data — only movements exceeding a threshold are flagged as potential falls and sent to the cloud. In the second collaboration stage, acceleration data is transmitted via TCP to a cloud server, where the FRESH algorithm extracts 794 time-series features. In the third cloud stage, the FEDT model — an ensemble of Classification and Regression Trees (CART) with a regularization term to prevent overfitting — makes the final fall/non-fall determination. If a fall is detected, an alert is sent back to the mobile device. This architecture splits the computational load: simple filtering happens on the resource-constrained watch, while complex classification runs on the cloud server.
Key findings
The FEDT algorithm outperformed both random forest and gradient boosted decision trees (GBDT) by 1-3% on both sensitivity and specificity across three public benchmark datasets (SisFall, MobiAct, MMsys) and one practical dataset collected by the researchers. On MobiAct, FEDT achieved 98.05% sensitivity and 99.95% specificity; on MMsys, 97.33% sensitivity and 99.97% specificity; on SisFall, 98.11% sensitivity and 99.98% specificity. The authors attribute FEDT's advantage over random forest to its regularization term that prevents overfitting, and over GBDT to its ensemble nature that avoids underfitting. Crucially, the system demonstrated cross-device robustness: when trained on Samsung watch data and tested on Huawei watch data (which produces numerically different accelerometer readings), FEDT maintained 93% sensitivity — important for real-world deployment where users may have different devices. The three-stage architecture effectively balances computational efficiency with detection accuracy, making real-time fall detection practical on consumer wearable hardware.
Relevance
Fall detection is a critical accessibility and independent living concern for older adults and people with mobility-related disabilities. Reliable automatic fall detection can enable people to age in place with greater safety, providing peace of mind that help will be summoned if they fall and cannot get up themselves. This paper's mobile cloud collaboration approach is notable for its practicality — it works with commercially available smartwatches rather than specialized medical devices, lowering the barrier to adoption. The cross-device robustness findings are particularly relevant for real-world deployment, though the 93% cross-device sensitivity leaves room for improvement. For accessibility practitioners, this work illustrates how machine learning and wearable technology can support independent living, though the reliance on cloud connectivity could be a limitation in areas with poor internet access. The system's privacy advantage over camera-based fall detection methods is also worth noting, as it only collects accelerometer data rather than images or video.
Tags: fall detection · machine learning · wearable technology · aging · health monitoring · cloud computing · ensemble learning