Non-intrusive fall detection monitoring for the elderly based on fuzzy logic

Highlights

• Accelerometer-based fall detection prompts to false fall detection.

• Fusion of sound sensor and accelerometer to increase the fall detection accuracy.

• A short burst sound in the range of 50–70 dB and duration less than 500 ms.

• Fuzzy logic-based fall detection algorithm.

• False fall detections per day decrease from high of 1.37 to low of 0.06.

Abstract

This paper presents a health condition monitoring solution that detects an elderly accidental fall occurrence. The fall detection algorithm implements both accelerometer-based and sound-based detections for the possible occurrence of a valid fall. The accelerometer-based fall detection is instrumental in the detection of a valid fall occurrence. However, it has been shown that by using accelerometer alone is insufficient to accurately detect a fall, as the accelerometer misinterprets some daily motion activities and classified them as valid falls. The sound sensor can be used to detect the sound pressure generated from a resultant fall, but sound pressure cannot by itself be used as a reliable indicator of a fall. Thus, a fuzzy logic-based fall detection algorithm is developed to process the output signals from the accelerometer and sound sensor, where a valid fall activity detected by the accelerometer, coupled with a detected sound pressure from the resultant fall can infer an occurrence of a valid fall. This paper demonstrates the fuzzy logic algorithm to improve the accuracy of detecting a valid fall as compared to the accelerometer only fall detection algorithm and it can be demonstrated that the algorithm is capable of minimizing false fall detections per day from high of $1.37$ to low of $0.06$.

Introduction

Falls are the major cause of both fatal and non-fatal injuries among people and create a hindrance in living independently. The frequency of falls increases with age and frailty level. Between 2007 and 2011, in Singapore, at least 50 elderly persons have been found dead in their own homes from causes relating to falls and illnesses [1]. With the rapid technological advancements, various small and non-intrusive remote health condition monitoring solutions have been proposed and developed with the objectives to solve or mitigate problems encountered by elderly people living alone, and ultimately to save lives by providing them with timely assistance. Commercial product developments and active academic research on fall detection have been motivated by the considerable risks of falls and the substantial increase of the elderly people population. A typical fall detection system has two major functional components: (a) the detection component, which detects falls and (b) the communication component that communicates with emergency contact after fall detection.

In Singapore, the government takes initiative in making elderly-friendly public housing so as to facilitate aging in place [2]. In 2013, a pilot condition monitoring project called Elderly Monitoring System (EMS) was deployed to 500 public housing flats occupied by lone elderly residents. These in-home condition monitoring and alert system monitors round the clock activity levels of each resident in a non-intrusive way, and trigger an alert to a designated caregiver in the event of anomalies [2]. With the initial success of the pilot runs, several similar health condition monitoring systems [3], [4] were also proposed and underwent trials by different competing solution providers aimed to solve or mitigate the same set of problems defined earlier.

Various health condition monitoring solutions proposed and demonstrated by different solutions providers, in many ways are similar to the condition monitoring idea where elderly people are monitored for motion activities. In most cases, optical camera and passive infrared (PIR) motion detectors are used for such purposes. The primary triggering criteria when a registered caregiver is alerted will be based on the abnormal lack of motion activities or from a manual trigger by an elderly person requesting for assistance.

With the best efforts to understand the implementation of the various condition monitoring solutions, an automatic and reliable method of detecting an elderly person falling down is currently missing or not actively promoted. This feature lapse is intentional, as the various methods for reliable fall detections are currently still in active research, and the available fall detection algorithms and methods are not able to provide $100\%$ human fall detection accuracy. A robust fall detection system is one that is able to classify the falls as “falls” and the non-falls as “non-falls” under real life conditions. If a fall event occurs and the system does not detect it, the consequences can be dramatic. In contrast, if the system reports an excessive number of false fall alerts, caregivers may perceive it as ineffective and useless, which may lead to device rejection.

There are commercially available systems that offer human fall detections, but these systems come with disclaimers stating that accuracy in detecting a valid human fall is not guaranteed. Several reviews [5], [6], [7], [8], [9] of the commercially available fall detection systems have shown that the commercially available systems are already available and deployed, but not in widespread use. The products are mainly offered as paid services for monitoring the safety of elderly people staying by themselves, and for eldercare centers. For the wearable products, they use either accelerometers or tilt sensors to detect a valid human fall.

To date, one of the most common implementation for detecting a fall requires an elderly person to wear a portable electronics wearable device with a built-in inertial sensor in the form of a tri-axial accelerometer, a wireless communication interface, and a battery. The accelerometer continuously detects motion accelerations in the three-dimensional vector space, and by analyzing the motion acceleration behavior, human fall occurrence can be ascertained or predicted. One of the well-known and practical accelerometer-based fall detection algorithm is developed by Ning Jia [10] using an Analog Devices ADXL345 digital MEMs tri-axial accelerometer [11]. The well-known algorithm detects a sequence of known motion-based activities (e.g., free-fall, weightlessness, strike, motionless and long time motionless) that can be pieced together in order to approximate a valid fall. In yet another well-known implementation, Bourke et al. [12], [13] developed a fall detection algorithm using a tri-axial accelerometer to detect fall impact and human posture. The algorithm, considered the sum of vectors of the accelerometer outputs and the detected posture to decide if a valid fall has occurred. Both algorithms are highly accurate in detecting a real human fall process. However, both algorithms are also sensitive to human motion attributed to daily movements (sitting, standing, etc.) and each human motion is person dependent. In both approaches, a change in body orientation from upright to lying that occurs immediately after a large negative acceleration indicates a fall. However, generally despite all the research dedicated to fall detection, there still does not exist a 100% reliable algorithm that catches all falls with no false alarms. Hence, both algorithms also provide unwanted and false positive human fall results. In field implementations, both algorithms suffer substantial setbacks in terms of the relatively large amount of false positive fall detections.

For each elderly person, individual movement and physical reaction to the occurrence of a fall is not the same [14], [15], [16], thus it is difficult for the algorithms to cater to all forms of fall patterns, hence the incurred setbacks of false fall detections. In order to have an accurate detection, both algorithms require the elderly person to physically move or react to a fall in a certain way expected by the device manufacturers, which is neither logical nor practical. Thus, using only accelerometer to detect a valid fall is insufficient when good accuracy with minimum false positives is desired.

In this paper, the authors propose an e-healthCM solution that automatically detects and predicts an elderly person accidental fall occurrence. The basic functionality of e-HealthCM is similar to the various health condition monitoring solutions for fall detection, where it monitors a senior citizen’s home for accidental fall activity, and to automatically request for assistance when a valid fall is detected. With reference to the discussed shortfalls and known restrictions of an accelerometer only fall detector, the e-HealthCM improves on the overall fall detection accuracy by providing a second level of sound-based fall sensing as an enhancement to the accelerometer only fall detector.

The remaining part of the paper is organized as follow: Section 2 presents the proposed hardware development. Section 3 presents the Fall detection Algorithm. Section 4 discusses the experiment and verification results and Section 5 concludes the paper.