RESNA Annual Conference - 2019

Evaluation Of Imu-Based Pre-Impact Fall Detection Algorithms Using Public Dataset

S. Ahn, B. Koo, H. Jeong, Y. Nam, J. Kim, T. Kim, Y. Kim

Department of Biomedical Engineering and Institute of Medical Engineering, Yonsei University


Fall is a significant cause of injury and death in the elderly [1]. The frequency of falls are increasing as the elderly population increases in many countries. Approximately 35% of community-dwelling elderly adults and 50% of those residing in long-term care facilities fall at least once per year [2,3]. Many of them suffer moderate to severe injuries that require hospitalization and increase the risk of death. Therefore, it is a major healthcare priority to develop fall prevention systems for the elderly adults.

Fall prevention strategies involve identifying individuals with an increased risk of falling and implementing the appropriate prevention mechanism. They include physical restraints [4], fall-related fracture prevention strategies [5-7], study of risk factors related to syncope [1] and multi-factorial risk assessment and management [8]. One strategy to prevent or reduce injury against falls is to detect them during descent (pre-impact fall detection) and to mitigate the impact [9-13]. Recently, a portable wearable sensor was used to measure acceleration and angular velocities during falls. If a fall can be detected in its earliest stage during descent, a more efficient impact reduction system can be implemented with a longer lead time. In this study, pre-impact fall detection algorithms were developed using IMU sensors placed on the subjects’ waists. The algorithms used subjects’ vertical angles and triangle features [14], which are features of tilting. The subjects engaged in four kinds of simulated falls and six types of ADLs to validate the pre-impact fall detection algorithms developed in this study. The performance of the algorithm using vertical angles (“VA”) and triangle feature (“TF”) was evaluated using the public dataset [15].


Figure 1. Simulated Falls (A) Forward fall (B) Backward fall (C) Lateral fall (D) Twist fall
Forty healthy male volunteers (age 23.4 ± 4.4 years, 68.7 ± 8.9 kg, 172.0 ± 7.1 cm) participated in the study. The experimental protocol was approved by the Yonsei University Research Ethics Committee (1041849-201308-BM-001-01), and written informed consent was obtained from each subject. In fall simulations, subjects were told to stand upright on the floor beside a soft foam mattress, then to fall (as if fainting) forward, backward (with and without a twist), or laterally (Figure 1). All falls were performed five times. A chair and mattress were used for ADL trials, which included sit-to-stand transitions, walking, stand-to-sit transitions, sit-to-lie transitions, jumping, and running. Each activity was performed three times.

Figure 2. Pre-impact fall detection algorithms (VA and TF algorithms)
An MPU-9150 motion-tracking device (InvenSense, San Diego, CA, USA) containing a 3-axis accelerometer and a 3-axis gyro sensor was used for pre-impact fall detection. The sensor was attached to the middle of the left and the right anterior superior iliac spines. Data were sampled at 100 Hz. All falls and ADLs were also recorded using a Bonita motion capture camera (Vicon Motion Systems Ltd., Oxford, UK) at 340 frames/s.

Data analysis was performed using MATLAB R2010a (MathWorks Inc., Natick, MA, USA). All data were low-pass filtered at 8 Hz. Acceleration data was transformed into the vertical angle in the sagittal and the frontal planes, measuring how many degrees these body segments deviated from the vertical axis (i.e., 0° is standing, and 90° supine on the floor). TF was defined by the area of the triangle consisting of the vector sum of the acceleration in the two directions (x-axes, z-axes) and the acceleration in the y-direction.

Table 1.  Lead times based on VA and TF algorithms.


VA Algorithm TF Algorithm
Forward Fall 403 ± 32.7 ms 423 ± 22.8 ms
Side Fall 422 ± 42.3 ms 422 ± 31.8 ms
Backward Fall 423 ± 33.1 ms 442 ± 47.4 ms
Twist Fall 381 ± 19.0 ms 397 ± 27.8 ms
Mean ± Std 401 ± 46.9 ms 427 ± 45.9 ms

The thresholds in the algorithm were optimized to maximize both the accuracy and the lead time. (Lead time was defined as the time between fall detection and impact.) Results showed that a fall was detected based on the VA algorithm when the vector sum of acceleration was less than 0.9 g, the angular velocity was greater than 47.3°/s, and VA was greater than 24.7° (Figure 2). Similarly, a fall was detected based on the TF algorithm when the vector sum of acceleration was less than 0.9 g, the angular velocity was greater than 47.3°/s, and TF was larger than 0.19 (Figure 2).


The pre-impact fall detection algorithms were tested for ten subjects. No failed detection occurred for four types of falls (100% sensitivity), and no incorrect detection was found for six different types of ADLs (100% specificity). Lead times for four different types of falls are shown in Table 1. Average lead time of the VA algorithm was 403 ± 32.7 ms, 422 ± 42.3 ms, 402 ± 33.1 ms, and 381 ± 19.0 ms for forward, lateral (side), backward, and twist falls, respectively. Average lead time of the FT algorithm were 423 ± 22.8 ms, 422 ± 31.8 ms, 442 ± 47.4 ms, and 397 ± 27.8 ms for forward, lateral (side), backward, and twist falls, respectively. The mean lead time of FT and VA algorithms were 427 ± 45.9 ms and 401.9 ± 46.9 ms respectively.

The two algorithms developed in this study were evaluated using the SisFall dataset [14]. Both algorithms successfully detected all falls, representing 100% sensitivity. VA algorithm and TF algorithm showed the specificity of 78.3% and 83.9%, respectively. Table 2 compares the performance of VA and TF algorithms with those from previous studies. TF algorithm showed 90.3% accuracy (100% sensitivity and 83.9% specificity), and VA algorithm showed 86.9% accuracy (100% sensitivity and 78.3% specificity).

Table 2.  Accuracy comparison with previous studies


Wu [11] Tamura et al. [16] Bourke et al. [9] This study
VA algorithm TF algorithm
Accuracy (%) 80.5 81.8 87.2 86.9 90.3
Sensitivity (%) 100 93 100 100 100
Specificity (%) 67.6 74.4 78.7 78.3 83.9
Feature Acc Acc, Gyro Vertical velocity Acc, Gyro, VA Acc, Gyro, TF

Discussion and Conclusions

In this study, pre-impact fall detection algorithms were implemented using an IMU sensor positioned at the waist. The algorithms used acceleration, angular velocity and one of the two tilting features (“VA” or “TF”).

Many studies have used pre-impact fall detection algorithms. Some studies have shown 100% specificity but without 100% sensitivity [9, 11]. In particular, those algorithms produced false-positive errors, mistaking jumps or stand-sit transitions for falls. If acceleration is used as the only threshold, jumping and sitting in a chair can be mistaken for falling. Our algorithms used tilting features (vertical angle or triangle feature) as a threshold in addition to acceleration and angular velocity in order to avoid such mistakes.

A previous study achieved a longer lead time of roughly 700 ms [10]. However, the algorithm required using two inertial sensors had lower accuracy. The algorithms developed in this study achieved 100% accuracy with only one sensor and our dataset. The results showed that the lead time was approximately 30ms longer in TF algorithm than in VA algorithm because TF increases nonlinearly with VA.

Although many fall detection algorithms have been developed, they have not been put to practical use because they were not developed using ADL and fall data from elderly subjects. Publicly available datasets were generated from a few activities, and none included fall data from elderly subjects. However, the SisFall dataset [14] was recently made publicly available and was developed based on data from more participants, ADLs, and falls than any other publicly available dataset. In addition, the SisFall dataset contained data from elderly people. In this study, the performance of two algorithms (VA algorithm and TF algorithm) was evaluated using the SisFall dataset [14]. VA algorithm should a specificity of 78.3% because it generated false positives for the seven types of ADLs. TF algorithm should a specificity of 83.9% because it generated false positives for the three types of ADLs. These ADLs involved transitions to lower positions, such as sitting in a low chair quickly, lying down quickly, or getting in a car.

In this study, pre-impact fall detection algorithms were developed using an inertial sensor unit. VA and TF algorithms detected falls with 100% accuracy with our dataset and resulted in lead times of 401 ± 46.9 ms and 427 ± 45.9 ms, respectively. In addition, the performance of the algorithms was evaluated using the SisFall dataset. Both algorithms detected every fall in the SisFall dataset (100% sensitivity). VA algorithm had a specificity of 78.3%, and TF algorithm had a specificity of 83.9%. They were shown to be more accurate than the existing algorithms. Improved fall detection would significantly improve the value of fall protection systems to protect the elderly from injuries sustained during falls.


[1] Kenny RA, O’Shea D. Falls and syncope in elderly patients. Clin Ceriatr Med. 2002 18: xiii-xiv.

[2] Nevitt MC, Cummings SR, Hudes ES. Risk factors for injurious falls: a prospective study. J gerontol. 1991 46(5): M164-70.

[3] Tinetti ME, Doucette J, Claus E., Marottoli R. Risk factors for serious injury during falls by older persons in the community. J Am Geriatr Soc. 1995 43(11): 1214-21.

[4] Gross YT, Shimamoto YC, Rose L., Frank B. Why do they fall? Monitoring risk factors in nursing homes. J Gerontol Nurs. 1990 16(6): 20-5.

[5] Smeesters C, Hayes WC, McMahon TA. Disturbance type and gait speed affect fall direction and impact location. J Biomech. 2001 34(3): 309-17.

[6] van den Kroonenberg AJ, Hayes WC, McMahon TA. Hip impact velocities and body configurations for voluntary falls from standing height. J Biomech. 1996 29(6): 807-11.

[7] Yamamoto S, Tanaka E, Ikeda T, Kubouchi Y, Harada A, Okuizumi H. Mechanical simulation for hip fracture by a fall using multibody-FE hybrid human model. 2006 J Biomech. 39(1): S89-90.

[8] Weatherall M. Multifactorial risk assessment and management programmes effectively prevent falls in the elderly. Evid base Healthc Publ Health, 2004 8(5): 270-2.

[9] Bourke AK, O’Donovan KJ, Laighin GÓ, A threshold-based fall-detection algorithm using a bi-axial gyroscope sensor. Med Eng Phys. 2008 30(1): 84-90.

[10] Nyan MN, Tay FE, Tan AW, Seah KH. Distinguishing fall activities from normal activities by angular rate characteristics and high-speed camera characterization. Med Eng Phys. 2006 28(8): 842-9.

[11] Wu G. Distinguishing fall activities from normal activities by velocity characteristics. J Biomech. 2000 33(11): 1497-500.

[12] Zhang T, Wang J, Xu L, Liu P. Fall detection by wearable sensor and one-class SVM algorithm. Lect Notes Control Inf Sci. Springer; 2006 345: 858–63.

[13] Ganti RK, Jayachandran P, Abdelzaher TF, Stankovic JA. Satire: a software architecture for smart attire. In: Proceedings of the 4th international conference on Mobile systems, applications and services. ACM; 2006 110–23.

[14] Ahn S, Choi D, Kim J, Kim S, Jeong Y, Jo M, Kim Y. Optimization of a Pre-impact Fall Detection Algorithm and Development of Hip Protection Airbag System. SENSOR MATER. 2018 30(8): 1743-52.

[15] Sucerquia A, López JD, Vargas-Bonilla JF. SisFall: A Fall and Movement Dataset. Sensors (Basel). 2017 17(1): 198.

[16] Tamura T, Yoshimura T, Sekine M, Uchida M, Tanaka O. A Wearable Airbag to Prevent Fall Injuries. IEEE Trans Inf Technol Biomed. 2009 13(6): 910-4.


This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (NRF-2018R1D1A1B07048575).