Phone based Fall Detection by Genetic Programming

Phone based Fall Detection by Genetic Programming Hoang Anh Dau, Flora D. Salim, Andy Song, Lachlan Hedin, Margaret Hamilton School of Computer Science and Information Technology, RMIT University Melbourne, Australia 3000

{dau.hoanganh, flora.salim, andy.song, lachlan.hedin, margaret.hamilton} @rmit.edu.au

ABSTRACT Elderly people are prone to fall due to the high rate of risk factors associated with ageing. Existing fall detection systems are mostly designed for a constrained environment, where various assumptions are applied. To overcome these drawbacks, we opt to use mobile phones with standard builtin sensors. Fall detection is performed on motion data collected by sensors in the phone alone. We use Genetic Programming (GP) to learn a classifier directly from raw sensor data. We compare the performance of GP with the popular approach of using threshold-based algorithm. The result shows that GP-evolved classifiers perform consistently well across different fall types and overall more reliable than the threshold-based.

Keywords Fall detection, genetic programming, mobile sensing

1.

INTRODUCTION

Automatic fall detection could enable timely rescue and follow-up to minimize possible negative impacts of the fall. Literature on fall detection research can be broadly divided into three categories, namely wearable device based, ambience sensor based and vision based. Almost all techniques require manually designing a feature set. With the ubiquitous uptake of smart phones, there has been increasing interest in exploiting their sensing capabilities. Most mainstream consumer phones now come with built-in sensors, originally for user interface enhancement. Various human activities can be inferred from the phone sensor data. We propose a learning based approach for fall detection. Our technique does not require manually setting thresholds or strenuously composing feature sets. The learning process is expected to find the discrimination rule to separate falls from non-falls (normal activities). The method is based on Genetic Programming (GP), a powerful evolutionary learning method. We enable GP to work directly on raw sen-

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit o r c ommercial a dvantage a nd t hat copies bear this notice and the full citation on the first p age. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MUM ’14 Melbourne Australia Copyright 2014 ACM 978-1-4503-3304-7/14/11 ...$15.00. http://dx.doi.org/10.1145/2677972.2678010

256

sor data of the accelerometer, gyroscope and magnetometer. Three fall types, specifically fall while standing, fall while walking and fall while sitting, are investigated. We perform experiments with two phone placement options, those are tight pants pocket and loose pants pocket.

2.

RESEARCH ON FALL DETECTION

Most of the studies on fall detection are based on thresholding data collected from one or more inertial sensors. The most common type of sensor used is accelerometer. Thresholdbased approaches are prone to high false alarms, especially when fall-like events are present. More sophisticated systems employ context-based reasoning methods to eliminate false positives. For example, location sensor data can be used to work out where the event happens [3]. This helps to differentiate a dramatic fall on the floor with a normal quick lay down on the sofa. Machine learning has been used for fall detection in place of manual thresholding [1] or in conjunction with thresholding to increase accuracy [4]. To the best of our knowledge, GP has not been used for detecting falls before.

3.

GP METHODOLOGY

GP is a global search technique inspired by biological evolution. The basic process of GP training involves generating an initial population of computer programs to solve the problem at hand. A fitness function is used to evaluate each program individual, and the better-performed programs are chosen to undergo genetic operation to create the next generation. This iterative process ceases when a terminating condition is met, in this case, a solution with 100% accuracy is found or the maximum number of generation is reached. Details of the GP representation used in this study can be found in our early work [6].

4.

EXPERIMENTS AND RESULTS

For each fall type, we collected 20 falls for training and 20 falls for testing. All falls are simulated by a male in his 20s with floor contact on a mattress. Between the falls, the subject performs other activities of daily life such as walking and running. Table 1 provides some statistics of the data collection. We sample the data at 10Hz and use sliding windows of size 12, overlapping by 1. These values are decided empirically. For comparison, we implemented the popular thresholdbased method similar to the one described in [2] and [5]. The threshold setting is based on the assumptions that the phone

has undergone a large acceleration and rotational change during a fall. Table 1: Data collection summary Number of subject 1

Table 2: Test result for HTC phone and Samsung phone for the tight pant pocket setting and loose pant pocket setting. We perform binary classification for each fall type (fall or non-fall). Mix-fall refers to the presence of all three fall types aforementioned in training. HTC-tight GP method Threshold Method

Sensing medium

Samsung Galaxy S4 GT-I9505 HTC One Accelerometer Gyroscope Magnetometer

Stand-fall

Sensor type

Loose front pant pocket Tight front pant pocket

Walk-fall

Sensor placement

Stand-fall: Fall while standing Walk-fall: Fall while walking Sit-fall: Fall from the chair

Sit-fall

Fall type

Mix-fall Table 2 displays the test result for the GP approach and threshold method. In terms of performance metrics, we report accuracy (Acc.), True Positive rate (TP) and True Negative rate (TN). We found the performance of the threshold technique deteriorates in the presence of sudden stop events. When the phone is placed in the loose pant pocket setting, which introduces noise as the phone can enjoy some degrees of movement, GP outperforms threshold method.

5.

CONCLUSION AND FUTURE WORK

We have shown that reliable fall detection can be achieved with GP. GP work directly on raw sensor input and does not rely on manually selected features. This means the requirement for human expert intervention can be minimised. For future works, we would like to investigate the reliability of our method on more fall scenarios and with more participating subjects.

6.

Acc. TP TN Acc. TP TN Acc. TP TN Acc. TP TN

93.75% 87.50% 100.00% 88.75% 77.50% 100.00% 61.25% 22.50% 100.00% 81.25% 62.05% 100.00%

GP method

Threshold Method

Stand-fall

Acc. TP TN Acc. TP TN Acc. TP TN Acc. TP TN

Acc. TP TN Acc. TP TN Acc. TP TN Acc. TP TN

Walk-fall

Sit-fall

Mix-fall

92.62% 83.22% 95.61% 86.19% 86.40% 86.16% 97.01% 98.14% 96.65% 88.19% 87.54% 88.39%

98.75% 97.50% 100.00% 92.50% 85.00% 100.00% 65.50% 30.00% 100.00% 85.42% 70.83% 100.00%

HTC-loose

GP method

Threshold Method

Stand-fall

Acc. TP TN Acc. TP TN Acc. TP TN Acc. TP TN

Acc. TP TN Acc. TP TN Acc. TP TN Acc. TP TN

Walk-fall

Sit-fall

Mix-fall

90.33% 84.77% 91.56% 95.19% 93.71% 95.32% 97.02% 96.86% 97.05% 75.36% 70.43% 76.12%

80.00% 67.50% 92.50% 71.25% 75.00% 67.50% 72.50% 50.00% 95.00% 74.58% 64.16% 85.00%

Samsung-loose

GP method

Threshold Method

Stand-fall

Acc. TP TN Acc. TP TN Acc. TP TN Acc. TP TN

Acc. TP TN Acc. TP TN Acc. TP TN Acc. TP TN

Walk-fall

Sit-fall

Mix-fall

257

92.59% 87.50% 96.17% 95.14% 94.24% 95.25% 97.79% 98.42% 97.64% 83.49% 80.08% 84.15%

Samsung-tight

REFERENCES

[1] M. V. Albert, K. Kording, M. Herrmann, and A. Jayaraman. Fall classification by machine learning using mobile phones. PloS one, 7(5):e36556, 2012. ˘ Zbrien, ´ [2] A. Bourke, J. Oˆ aA and G. Lyons. Evaluation of a threshold-based tri-axial accelerometer fall detection algorithm. Gait & posture, 26(2):194–199, 2007. [3] H. Gjoreski, M. Luˇstrek, and M. Gams. Context-based fall detection using inertial and location sensors. In Ambient Intelligence, pages 1–16. Springer, 2012. [4] S. Kozina, H. Gjoreski, M. Gams, and M. Luˇstrek. Efficient activity recognition and fall detection using accelerometers. In Evaluating AAL Systems Through Competitive Benchmarking, pages 13–23. Springer, 2013. [5] F. Sposaro and G. Tyson. ifall: an android application for fall monitoring and response. In Engineering in Medicine and Biology Society, 2009. EMBC 2009. Annual International Conference of the IEEE, pages 6119–6122. IEEE, 2009. [6] F. Xie, A. Song, and V. Ciesielski. Genetic programming based activity recognition on a smartphone sensory data benchmark. In Evolutionary Computation (CEC), 2014 IEEE Congress on, pages 2917–2924. IEEE, 2014.

Acc. TP TN Acc. TP TN Acc. TP TN Acc. TP TN

92.59% 90.64% 93.16% 88.98% 91.78% 88.63% 97.36% 97.59% 97.32% 78.50% 89.32% 76.50%

83.75% 79.00% 97.50% 77.50% 80.00% 75.50% 73.75% 47.50% 100.00% 78.33% 65.83% 90.83%