Reducing computational cost for real-time fall detection: A Mahalanobis pre-filtering approach
DOI:
https://doi.org/10.54939/1859-1043.j.mst.114.2026.47-54Keywords:
Fall detection; Mahalanobis distance; TinyML; Wearable sensors; Computational efficiency.Abstract
Real-time fall detection using wearable devices is essential for TinyML applications, but continuous neural network inference can be expensive for low-power microcontrollers. This paper presents a two-stage fall detection framework that utilizes Mahalanobis distance as a lightweight pre-filter before a CNN classifier. In the first stage, accelerometer data is converted into a nine-dimensional statistical feature vector and compared against a reference distribution of non-fall movements. Windows that fall below a specified threshold are considered low-risk and bypass the CNN, while suspicious ones proceed to the second stage for classification. Evaluated on the KFall dataset with 62,204 windows, the Mahalanobis pre-filter achieved fall recall rates of 98.18% and 98.55%, and false-positive rates of 39.35% and 37.40%, respectively. On an ESP32-S3 microcontroller, the first stage required only 196 microseconds per window, resulting in an average computation reduction of about 58.9-60.6% compared to CNN-only inference.
References
[1]. Z. Jiang et al., “Fall Detection Systems for Internet of Medical Things Based on Wearable Sensors: A Review”, IEEE Internet Things J., vol. 11, no. 21, pp. 34797–34810, (2024). DOI: 10.1109/JIOT.2024.3421336.
[2]. A. Ramachandran and A. Karuppiah, “A Survey on Recent Advances in Wearable Fall Detection Systems”, BioMed Research International, vol. 2020, no. 1, p. 2167160, (2020). DOI: 10.1155/2020/2167160.
[3]. J. Liu et al., “A review of wearable sensors based fall-related recognition systems”, Engineering Applications of Artificial Intelligence, vol. 121, p. 105993, (2023). DOI: 10.1016/j.engappai.2023.105993.
[4]. C. V. F. Pereira, E. M. De Oliveira, and A. D. De Souza, “Machine Learning Applied to Edge Computing and Wearable Devices for Healthcare: Systematic Mapping of the Literature”, Sensors, vol. 24, no. 19, p. 6322, (2024). DOI: 10.3390/s24196322.
[5]. H. Haresamudram, C. I. Tang, S. Suh, P. Lukowicz, and T. Plötz, “Past, Present, and Future of Sensor-based Human Activity Recognition Using Wearables: A Surveying Tutorial on a Still Challenging Task”, Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 9, no. 2, pp. 1–44, (2025). DOI: 10.1145/3729467.
[6]. O. Aziz, M. Musngi, E. J. Park, G. Mori, and S. N. Robinovitch, “A comparison of accuracy of fall detection algorithms (threshold-based vs. machine learning) using waist-mounted tri-axial accelerometer signals from a comprehensive set of falls and non-fall trials”, Med Biol Eng Comput, vol. 55, no. 1, pp. 45–55, (2017). DOI: 10.1007/s11517-016-1504-y.
[7]. J. He, Z. Zhang, X. Wang, and S. Yang, “A Low Power Fall Sensing Technology Based on FD-CNN”, IEEE Sensors J., vol. 19, no. 13, pp. 5110–5118, (2019). DOI: 10.1109/JSEN.2019.2903482.
[8]. X. Yu et al., “A practical wearable fall detection system based on tiny convolutional neural networks”, Biomedical Signal Processing and Control, vol. 86, p. 105325, (2023). DOI: 10.1016/j.bspc.2023.105325.
[9]. J. Peng, X. Zhang, and H. Li, “A LSTM-Based Fall Prediction Method Using IMU”, Advances in Asian Mechanism and Machine Science, N. V. Khang, N. Q. Hoang, and M. Ceccarelli, Eds., Mechanisms and Machine Science, vol. 113, Cham: Springer International Publishing, pp. 680–690, (2022). DOI: 10.1007/978-3-030-91892-7_65.
[10]. X. Hu, S. Yu, J. Zheng, Z. Fang, Z. Zhao, and X. Qu, “A hybrid CNN-LSTM model for involuntary fall detection using wrist-worn sensors”, Advanced Engineering Informatics, vol. 65, p. 103178, (2025). DOI: 10.1016/j.aei.2025.103178.
[11]. T.-H. Nguyen et al., “A frame-level video annotation tool for dynamic gestures and apply in Vietnamese sign language”, J. Mil. Sci. Technol., vol. 112, pp. 167–175, (2026). DOI: 10.54939/1859-1043.j.mst.112.2026.167-175.
[12]. Y. Li et al., “A Decade of Progress in Wearable Sensors for Fall Detection (2015–2024): A Network-Based Visualization Review”, Sensors, vol. 25, no. 7, p. 2205, (2025). DOI: 10.3390/s25072205.
[13]. S. R. Shakya, E. Ceh-Varela, J. Martinez, M. Fisher, and H. Allamezadeh, “Optimized CNN-Based Fall Detection for Elderly Using Wearable Devices on Resource-Constrained Devices”, 2025 IEEE 18th Dallas Circuits and Systems Conference (DCAS), Arlington, TX, USA: IEEE, pp. 1–6, (2025). DOI: 10.1109/DCAS65331.2025.11045488.
[14]. V. Divya and R. L. Sri, “Docker-Based Intelligent Fall Detection Using Edge-Fog Cloud Infrastructure”, IEEE Internet Things J., vol. 8, no. 10, pp. 8133–8144, (2021). DOI: 10.1109/JIOT.2020.3042502.
[15]. T. Sözer, “A cost-effective embeddable algorithm for accelerometer-based fall detector”, Engineering Science and Technology, an International Journal, (2025). DOI: 10.1016/j.jestch.2025.102185.
[16]. N. B. Nguyen, Assoc. Prof. D. N. Pham Minh, A. Phan Huy, and Pham Hoang Hung, “Enhance micro-Doppler signatures-based human activity classification accuracy of FMCW radar using the threshold method”, J. Mil. Sci. Technol., vol. 95, pp. 20–28, (2024). DOI: 10.54939/1859-1043.j.mst.95.2024.20-28.
[17]. G. J. McLachlan, “Mahalanobis distance”, Reson, vol. 4, no. 6, pp. 20–26, (1999). DOI: 10.1007/BF02834632.
[18]. X. Yu, J. Jang, and S. Xiong, “A Large-Scale Open Motion Dataset (KFall) and Benchmark Algorithms for Detecting Pre-impact Fall of the Elderly Using Wearable Inertial Sensors”, Front. Aging Neurosci., vol. 13, p. 692865, (2021). DOI: 10.3389/fnagi.2021.692865.
