In the last decade, many groups have carried out research and development on wearable electronics and sensors for unobtrusive, ambulatory and daily-life monitoring of human subjects. The results obtained have shown the possibility to use personal wearable devices to assist and support chronic patients [1
], elderly people [8
], emergency operators [10
] and also healthy subjects for sports, wellness and prevention [12
]. At the same time, the wearable technology market has exploded and is expected to further increase over the next few years, as proved by the growing interest of big players such as Google, Apple and Samsung.
The current trend is to augment objects worn on the body—e.g., watches, glasses, bracelets—with information and communications technology (ICT) to enable a bi-directional data exchange with a smartphone. These wearable devices or simply wearables
have been initially conceived as technological gadgets
but have the potential to support the user in the self-management of his/her health and wellness. Indeed, smart bracelets and/or watches can include physiological (e.g., photoplethysmography, electrodermal activity) or inertial (accelerometers, gyroscopes) sensors able to perform real-time monitoring of subject’s health parameters and movement/physical activity. Recent studies have reported the first attempts to employ smart watches/bracelets in e-health applications [15
] and many more are expected in the years to come.
Another trend, less explored but not less promising, is to embed ICT devices inside the shoe. The shoe is the ideal place to integrate sensors and communications technology: it has enough space for the micro-devices and it is the object that every one wears for most of the day. This last aspect is the key factor to enable user’s acceptance, as the user does not have to wear additional items and the technology can be completely hidden and transparent for him/her. The integration of inertial and force sensors in the shoe may enable a wide number of applications, ranging from simple activity/fitness tracking (e.g., activity classification, step count, burned calories) fragile people assistance (e.g., fall detection, pedestrian navigation) to complex biomedical assessment and gait analysis.
is the study of human locomotion, and it is used to assess and treat patients with conditions affecting their walking activity [20
]. The way we walk consists of consecutive gait cycles
. Each gait cycle includes a predefined sequence of phases (Heel-strike HS, Stance ST, Heel-off HO, Swing SW; see Appendix A
for a reference on the adopted terminology). Both temporal
gait parameters are important to assess a disease and/or a traumatic event, and also to define and optimize the treatment (e.g., rehabilitation, physical therapy). The temporal and spatial gait parameters are used in many biomedical and e-health applications, such as assessment of the recovery in stroke patients [21
] and gait-cycle-based control of functional electrical stimulators (FES) for drop foot
]. In the robotic rehabilitation field, the quantitative evaluation of the gait parameters allows the quantification of the improvements of the gait patterns [26
]. In addition, the spatial gait parameters, such as stride length, can be associated with fall risk [27
], or elaborated for foot motion localization in applications such as emergency operator rescue and pedestrian navigation [28
The reviews from Rueterbories et al. [32
] and from Taborri et al. [33
] provide a full overview of the gait phase detection methods and technologies. Many research works performed gait phase detection through inertial sensors (accelerometers, gyroscopes, inertial measurement units - IMU) applied to different body segments (pelvis, thigh, shank, foot) [34
]. As a recent relevant example, the work of Van Nguyen et al. [39
] focuses on an IMU sensor for an accurate estimation of foot position, velocity and attitude. Many other works were focused on on-foot
sensors. Most of the on-foot
systems dealt with force based methods, employing foot-switches or force sensitive resistors (FSRs) to measure the body/ground force interaction [40
]. Force based methods have reliable performance, but are unable to discriminate walking activities from load changes (i.e., in foot drop control, this implies the user turning off the detection/control system at the end of the walking activity). To solve this issue, Pappas et al. [43
] combined force and inertial sensors to obtain a reliable gait phase detection system, able to discriminate walking activities from load shifting in static tasks. They employed three FSRs (one under the heel and two under the fore-foot region) and a gyroscope attached to the back side of the shoe (above the heel). Force sensors detected the foot loading/unloading, while the gyroscope estimated the foot inclination and rotational velocity. In their work, Pappas et al. [43
] detected four gait phases through a state machine whose transitions were governed by a rule based algorithm applying predefined thresholds on the parameters extracted from the sensors (foot loading/unloading, inclination, rotational velocity). Other examples of force and inertial sensors combination can be found in [44
]. As underlined in [33
], gait phase partitioning algorithms can be divided into three main classes. Threshold-based
methods—such as the above reported from Pappas et al. [43
]—which apply predefined or adaptive thresholds to the sensor signals, are simple and are often suitable for integration in embedded systems. Machine learning
methods and in particular Hidden Markov Models (HMM) are more complex than threshold based methods but have shown improved performance in gait phase detection. As a relevant example, Mannini and Sabatini [37
] developed an HMM classifier which detected four gait phases from an uni-axial foot-mounted gyroscope, and achieved significantly better performace than the threshold-based method applied to the same dataset. A more recent trend is to apply hierarchical decision to the output of two or more HMMs. As reported in the work from Taborri et al. [38
methods provide excellent performance and are compatible with real-time implementation.
In-shoe sensor systems have been commonly developed for real-time detection of gait parameters and walking patterns with applications in the assessment of specific foot pathologies (e.g., flat foot [46
], diabetic foot [47
]) and posture/activity recognition for healthy subjects [48
] or people affected by neurological conditions such as stroke [51
] or celebral palsy [26
]. In-shoe sensor systems are generally based on movable pressure sensing insoles and/or inertial sensors, combined with an external electronic module (signal acquisition/pre-elaboration, data transmission). Examples of in-shoe systems can be found in the works from Edgar et al. [41
] and Bae et al. [54
Commercial products are limited to professional instruments for clinical evaluation and to a few consumer devices for sports and training of healthy users. Professional products include the F-scan [55
] (Tekscan Inc., Boston, MA, USA) and the Pedar [56
] (Novel Inc., Munich, Germany) systems that are sensing insoles for the monitoring of dynamic temporal and spatial pressure distributions. Example applications of professional products are gait stability analysis [57
], gait phase detection [58
] and analysis of the gait characteristics during running [59
]. Despite the reliable performance and the high spatial resolution, the professional systems are not suitable for long-term monitoring in daily life conditions: both systems are expensive (on the order of several kEuros) and use electrical wires to connect the insole to the waist-worn acquisition system. The consumer products are generally made by applying an external measurement and transmission unit to a dedicated shoe (e.g., Adidas miCoach
, Nike + iPod
). The common aspect is the reduced number of sensors (typically only one inertial sensor) and the interaction with the smartphone in which a dedicated app can deliver special information to the users (e.g., workout time, velocity, distance travelled, calories), engaging them for reaching higher performance during physical activity.
In the current paper, we assessed a recently developed commercial smart shoe for automatic gait phase detection in level walking. The prototype we employed is the FootMoov
smart shoe [60
] produced by the Italian shoe factory Carlos srl (Fucecchio, Firenze, Italy). FootMoov was originally designed as a mobile game controller and for simple physical activity training and coaching. FootMoov can be interfaced with the smartphone through a WiFi connection and has built-in force sensors (heel and forefoot), triaxial accelerometer (forefoot), chargeable battery and acquisition/transmission module. We chose to assess and develop the gait phase detection algorithm for FootMoov since, unlike the current professional and consumer products, all of the hardware—including sensors and electronics—is integrated inside
the shoe, and no external modules or application of additional parts are needed.
On the other side, FootMoov is a consumer product. Thus the sensor number and locations are not optimized for gait analysis. In addition, the precise sensor locations and orientation inside the shoe is unknown. In accordance with the current trends in wearable technology, the aim of this work is to show the possibility to employ a low cost smart shoe as a tool for biomedical and e-health applications, allowing the continuous and long-term monitoring of users/patients in daily life. To the best of our knowledge, the assessment of such a wearable product does not exist in the current literature.
In particular, we assessed the smart shoe for the real-time quantification of the temporal parameters of gait. We developed a dedicated gait phase detection algorithm that combines the information extracted from the force and accelerometer sensors of the FootMoov smart shoe. In a first experiment, we collected data on ten healthy subjects in free level walking conditions to perform a preliminary and qualitative assessment of the system and algorithm performance. In this test, the algorithm recognized 5925 strides over the total amount of 6000 strides (98.7%). In a second test, we performed a quantitative evaluation of the system in comparison with a reference gait phase signal obtained by an optical motion capture instrument. We evaluated the time difference in the onset of the detected gait phases with respect to the reference (mean error of 44.7 ms) and the error in the estimation of the single phase durations (minimum error of 0.036 s for Heel Strike , with a maximum error of 0.11 s for Heel Off).