The experimental Ocha-House is in the Bunkyo district in the central area of the Tokyo metropolitan region. The project was originally designed as a ubiquitous computing house that allows the mounting of various sensing devices [25]. According to Japanese standards, the Ocha-House corresponds to an extended 1LDK dwelling, that is, a 1-bedroom apartment with a kitchen separated from the living and dining areas. An overview of the building and experimental area is shown in Figure 1.

The experimental Ocha-House is in the Bunkyo district in the central area of the Tokyo metropolitan region (Figure 1A). The total experimental surface area was 42 m² (Figure 1B). A total of 8 high-sensitivity uniaxial accelerometers (shear-type pickup PV-87; Rion Co. Ltd) were installed on the floor to measure the floor vibrations occurring on the experimental surface. The PV-87 sensor characteristics were specified as follows by the manufacturer: charge sensitivity +40 pC to –40 pC/m/s²; range of detection 1 to 3000 Hz; dimensions 24 mm (hex) × 30.5 (H) mm; mass 115 g. The optimal number of sensor units and their locations were determined through a series of preliminary experiments. These experiments involved progressively increasing the sensitivity setting of the sensors and the number of units placed on both the west and east sides of the Ocha-House. The operation continued until the coverage was deemed sufficient to detect human motion across all parts of the experimental surface. The data related to these preliminary experiments are not presented here. The sensors were mounted on the floor using double-sided tape, as recommended by the manufacturer, and connected to 4 UV-16 2-channel charge amplifiers (Rion Co Ltd) configured in accordance with the manufacturer’s recommendations. The floor-vibration data acquisition was performed using USB-6008 data acquisition devices (National Instrument Corp), a laptop equipped with MATLAB 2015b, and the necessary data acquisition toolbox (MathWorks Inc). The signal was digitized at a 100-Hz sampling rate with a resolution of 12 bits. The abovementioned system is described hereafter as the floor-vibration monitoring system.

In total, 10 female participants engaged in 4 activities in the Ocha-House. Participants were recruited from the Ochanomizu University campus, which is a women’s university. They were selected based on the inclusion criterion of being aged ≥18 years, with the exclusion criterion being physical imbalance. Participant characteristics are summarized in Table 1, and the details of the 4 activities performed at the Ocha-House are presented in Figure 2. Before the commencement of the experiment, each participant completed a brief walking trial in the Ocha-House, lasting approximately 1 minute. In total, 2 researchers with expertise in gait analysis visually determined the gait type of each participant, specifically identifying whether they exhibited a heel strike or a lighter mid-strike or forefoot-strike landing. The walking trial revealed that all participants could be categorized into either heel-strike landing or mid-strike or forefoot-strike landing categories, with no other gait types observed. All experiments were conducted without any footwear, including slippers. In total, 9 participants wore socks. Moreover, 1 participant did not wear socks on the day of the experiment and completed the protocol barefoot.

In total, 4 activities were selected from the “inactivity quiet/light” or “home activities” categories of the compendium of physical activities [26]. The metabolic equivalent of task (MET) values, indicating the intensity of each activity, were reported in the compendium as follows:

Each activity lasted 6 minutes. To balance the contribution of the 3 tasks within “cooking, setting the table, and serving food” and considering the volume of the Douglas bag used for energy expenditure measurement (see the section Indirect Calorimetry), participants were orally given time information. This ensured that they spent approximately 40 seconds on each task every 2 minutes. The floor-vibration monitoring system recorded floor vibrations, estimate the number of steps, and compute quantitative parameters as described in the Floor Vibration Signal Treatment and Data Feature Extraction section. The participants wore 2 ActiGraph GT9X monitors (ActiGraph LLC) at the waist and wrist. The 10-second epoch activity count and the step-count prediction were recorded for each activity. Finally, Douglas bags were used during the last 2 minutes of each activity to collect the air expired by the participants, perform indirect calorimetry measurements, and obtain the actual energy expenditure. Throughout the experiments, the researchers stood quietly on an insulated part of the floor outside the experimental area to avoid producing confounding vibrations. The experiments were video recorded (data not shown).

The experimental protocol was approved by the Ochanomizu University Research Ethics Committee (2018-18). All the participants provided written informed consent and they did not receive any compensation. The individual shown in Figure 2 provided informed consent for the publication of their image.

The 8-sensor floor-vibration data of each 6-minute activity corresponded to 8 time series of 36,000 samples. Raw data were expressed in volt. For each series, the floor-vibration signal was rectified and smoothed using a Butterworth filter. The vector norms of the 8 sensors were then computed. A location-based calibration coefficient was applied to the vector norm for each data sample to uniformize the vibration magnitudes throughout the experimental surface (Multimedia Appendix 1). The 3 following data features were extracted:

The signal treatment and computation of the 3 previously described parameters were performed using the toolboxes included in the SciPy library [28]. An overview of the entire data processing process is shown in Figure 3.

The participants were equipped with 2 ActiGraph GT9X monitors worn on the waist and the wrist. The waist-worn device was positioned near the right hip of the participant, on the belt, or on the upper edge of the skirt or the bottom of the trousers. An elastic belt provided by the manufacturer was used when the clothes worn by the participant did not allow the tight mounting of the monitor. The wrist-worn GT9X device was mounted tightly on the nondominant hand always in the same direction. The 2 monitors were mounted by the same experimenter for all participants. The ActiGraph monitor data were collected in 1-second epochs. The 10-second epoch activity count data were extracted from the wrist-worn as well as waist-worn devices, and the “Crouter adult (2010)” equation was used to compute energy expenditure predictions (ActiLife 6; ActiGraph LLC) [29]. This algorithm uses a refined 2-regression model to distinguish between walking and lifestyle activities. In this study, the activity intensities were expressed in MET (ie, energy expenditure/participant weight/6 min). The number of steps estimated by the waist- and wrist-worn monitors was also recorded.

The actual energy expenditure was measured using the Douglas bag method during the last 2 minutes of each activity. The air composition of the bags was analyzed using a mass spectrometry gas analyzer (ARCO-1000; Arco System) calibrated on each experimental day in accordance with the manufacturer’s instructions. The gas volume was determined using a gas meter (DC-5; Shinagawa). The energy expenditure was estimated from oxygen consumption (VO₂) and carbon dioxide production (VCO₂) using the Weir formula (ie, 3.9VO₂+1.1VCO₂). In addition, the resting metabolic rate of each participant was evaluated before the experiment. The participant lay for 15 minutes on the bed of the Ocha-House, and the expired air was collected during the last 2 minutes. The actual MET value for each activity was calculated as the activity energy expenditure divided by the resting metabolic rate.

The experimental sessions were video recorded using an Arrows M03 smartphone (Fujitsu Ltd) or an iPad Mini 3 (Apple Inc). In total, 2 independent investigators inspected the videos and counted the number of actual steps for the 4 activities of each participant. In this study, a “step” is defined as the shift of the body weight support from one leg to the other, which includes a single-leg support phase and occurs at least partially on the anterior-posterior axis.

The 4 activities were compared for the floor count, step count, moving distance parameters using an ANOVA or the Kruskal-Wallis test, and multiple comparison procedures (Tukey or Nemenyi tests) were performed to locate differences. The same analysis was conducted for parameters extracted from the waist- and wrist-worn GT9X monitors and for the indirect calorimetry measurements. The relationships between the actual intensities measured by indirect calorimetry and floor count, step count, and moving distance were investigated using single linear regression tests. Multiple regression models were used to explore the relationship between different combinations of descriptors, including floor count, step count, and moving distance, and the actual intensities measured using indirect calorimetry. This analysis was conducted hierarchically. First, the regressions were only performed on the floor-vibration–extracted parameters. Second, the participant characteristics (ie, body weight and gait type) were included in the model descriptors. Additional hierarchical models are presented in Multimedia Appendix 2. The model with the highest R² value best explained the variation in the data and was selected for subsequent testing. Finally, mixed model ANOVA and post hoc pairwise operations were used to compare the performance of the best models built on data obtained with the floor-vibration–based monitoring system, the waist- and wrist-worn GT9X monitors, against the actual intensity and the actual number of steps.

The underlying assumptions for each test were evaluated before conducting the analyses. Data are presented as mean (SD). The statistical analysis was performed using the following Python libraries: StatsModels (0.13.2), Pingouin (0.5.3), and Scikit-Posthocs (0.7.0) [30,31]. Data used for the statistical analysis are shared in the Multimedia Appendix 3.

Activity intensities calculated from indirect calorimetry measurements were as follows: 1.2 (SD 0.2) MET for the sitting behavior, 1.9 (SD 0.4) MET for the ironing activity, 2.5 (SD 0.4) MET for the cooking activity, and 3.7 (SD 0.6) MET for the cleaning activity (Figure 4A). Pairwise comparisons indicated significant differences in intensity between sitting and cooking, between sitting and cleaning, and between ironing and cleaning (P=.004, P<.001, and P=.01, respectively). The actual number of steps was 0 (SD 0) for sitting, 48 (SD 36) for ironing, 133 (SD 35) for cooking, and 281 (SD 48) for cleaning (Figure 4B). Pairwise comparisons indicated significant differences in steps between sitting and cooking, between sitting and cleaning, and between ironing and cleaning (P=.003, P<.001, and P=.003, respectively).

The floor count parameter (arbitrary units) was as follows for each activity: 113 (SD 58) for sitting, 610 (SD 277) for ironing, 1046 (SD 748) for cooking, and 2323 (SD 1255) for cleaning the room (Figure 4E). Pairwise comparisons indicated significant differences in intensity between sitting and cooking, between sitting and cleaning, and between ironing and cleaning (P=.006, P<.001, and P=.04, respectively). The estimated number of steps was as follows: 12 (SD 7.3) for sitting, 78 (SD 27) for ironing, 133 (SD 53) for cooking, and 251 (SD 59) for cleaning (Figure 4F). Pairwise comparisons indicated that the respective intensities of the 4 activities were all significantly different (between sitting and ironing, P=.01; between ironing and cooking, P=.04; and others, P<.001). Finally, the moving distance parameter scored as follows: 1.3 (SD 1.8) m for sitting, 17 (SD 11) for ironing, 30 (SD 29) for cooking, and 115 (SD 39) for cleaning (Figure 4G). Pairwise comparisons indicated significant differences in intensity between sitting and cooking, between sitting and cleaning, and between ironing and cleaning (P=.03, P<.001, and P=.03, respectively). Furthermore, a significant correlation was found between the moving distance outcomes computed from the floor-vibration signal and the moving distances estimated from the video records (Figure S1 in Multimedia Appendix 2).

The waist-worn activity tracker estimated the activity intensities as follows: 1.1 (SD 0.2) MET for sitting, 1.2 (SD 0.3) MET for ironing, 2.1 (SD 0.4) MET for cooking, and 4.2 (SD 0.7) MET for cleaning (Figure 4C). Pairwise comparison analyses indicated significant differences between sitting and cooking, between sitting and cleaning, and between ironing and cleaning (P=.01, P<.001, and P<.001, respectively). The estimated number of steps were as follows: 0.9 (SD 2.4) for sitting, 8 (SD 12) for ironing, 55 (SD 27) for cooking, and 165 (SD 51) for cleaning (Multimedia Appendix 2). Pairwise comparison analyses indicated significant differences between sitting and cooking, between sitting and cleaning, and between ironing and cleaning (P=.007, P<.001, and P<.001, respectively).

The wrist-worn activity tracker estimated the activity intensities as follows: 1.8 (SD 0.4) MET for sitting, 6.4 (SD 0.5) MET for ironing, 5.6 (SD 0.7) MET for cooking, and 6.8 (SD 0.4) MET for cleaning (Figure 4D). Pairwise comparisons indicated significant differences between sitting and ironing and between sitting and cleaning (P=.002 and P<.001, respectively). The estimated number of steps was as follows: 11 (SD 5.7) for sitting, 177 (SD 39) for ironing, 131 (SD 37) for cooking, and 197 (SD 35) for cleaning (Multimedia Appendix 2). Pairwise comparison analyses indicated significant differences between sitting and ironing and between sitting and cleaning (P<.001 for both).

The results for the activity count parameters of the waist- and wrist-worn devices are shown in Multimedia Appendix 2.

As presented in Table 2, floor count, step count, and moving distance were significantly associated with the intensity of physical behavior (r²=.56, r²=.82, r²=.66, respectively; P<.001). Combining the 3 parameters resulted in an r² value of 0.84. Combining floor count, step count, and moving distance with participant personal characteristics, such as body weight and gait type, allowed predicting the intensity of physical behaviors with an accuracy of 88% (Table 1). The results of the additional hierarchical models are presented in Multimedia Appendix 2.

The mixed model ANOVA showed a significant effect of the measurement method and a significant interaction with the activity for both the predictions of the number of steps and activity intensities (P<.001).

Significant underestimations in the number of steps were noted for the waist-worn ActiGraph device predictions when all activities were considered together (Figure 5A, “total”). Regarding the prediction of activity intensities (Figure 5B), the pairwise comparison analyses revealed statistically significant overestimations for the wrist-worn ActiGraph device across all activities, The wrist-worn device and the floor-vibration system did not exhibit any difference with the actual number of steps. When activities are considered separately, pairwise comparisons indicate that the wrist-worn ActiGraph device underestimated the number of steps completed during cleaning but overestimated it for sitting and ironing activities. The floor-vibration–based predictions overestimated the number of steps required for sitting.

Regarding the prediction of activity intensities, the pairwise comparison analyses revealed statistically significant overestimations for the wrist-worn ActiGraph device across all activities. The floor-vibration system did not show any differences with the actual intensities evaluated by indirect calorimetry. Finally, the waist-worn ActiGraph device showed a slight but significant underestimation for the estimations of the ironing activity intensities. The results are shown in Figure 5.

This study presented a novel quantitative method that uses the monitoring of floor vibrations for the evaluation of physical behaviors at home. The floor count, step count, and moving distance parameters were computed from the floor-vibration signal. Statistical models combining these 3 parameters showed significant correlations with the actual energy expenditure measured in 10 participants in a structured experiment that included 4 common home-based activities. In addition, the step count parameter did not show any significant difference with the number of steps completed by the participants during the experiment. The predictions of the floor-vibration monitoring system for both the activity intensity and number of steps were equal to or more accurate than those obtained by the Actigraphy method using the refined 2-regression model.

The actual activity intensities measured by the indirect calorimetry for the 4 activities increased in accordance with the intensities presented in the compendium of physical activities [26], that is, 1.2 versus 1.3 MET; 1.9 versus 1.8 to 2.0 MET; 2.5 versus 2.5 MET; and 3.7 versus 3.3 MET for sitting (and watching video), ironing (and folding clothes), cooking (and setting the table), and cleaning the room, respectively (Figure 4A). The floor count, step count, and moving distance parameters also increased gradually for the 4 activities, indicating the feasibility of evaluating the intensity of physical behaviors in the home environment using the information provided by floor vibrations (Figures 4E-4G). Among them, step count was the only parameter that showed significant differences between all 4 activities. The single regression analyses also indicated a strong association between the step count and the actual activity intensities (r²=0.82; P<.001), whereas floor count and moving distance showed a weaker but still significant association with the actual activity intensities (r²=0.56 and 0.66, respectively; P<.001 for both). These pieces of information taken together may suggest that the estimated number of steps extracted from the floor-vibration signal could be used to make a reliable quantitative estimation of physical behavior in home settings. All the 3 parameters were designed to describe the inhabitant’s motion. Although the step count and floor count parameters can capture the physical dimension of the movement, the moving distance parameter adds a spatial dimension to the evaluation. The better performance of step count alone compared with moving distance alone may be due to the location approximation inherent to the limited number of sensors used to cover the entire house surface (Multimedia Appendix 2). A step is detected and not affected by any approximation. On the other hand, floor count may also be susceptible to inaccuracies, potentially due to variations in floor-vibration wave attenuation. Factors influencing this attenuation include the proximity of furniture, its weight and contact surface with the floor, proximity of support beams and walls, and the irregular geometry of the Ocha-House floor area.

Despite the possible weaknesses of the floor count and moving distance parameters, multiple linear regression models combining step count, floor count, and moving distance still exhibited stronger associations (r²≥0.85; Table 1; Table S1 in Multimedia Appendix 2). As presented in Table 2, the inclusion of body weight and gait type parameters as descriptors greatly lowered the contribution of floor count to the model. Indeed, although floor count does not correlate with either body weight or gait type (Table S2 in Multimedia Appendix 2), it is still the only parameter extracted from floor vibrations that can quantitatively capture the forces applied on the floor. Given that the actual body weight can be easily inputted into any smart home system, the question of the relevance of extracting and using the floor count parameter to make accurate predictions of energy expenditure remains open. Additional studies, including a population with more heterogeneous anthropometric characteristics, may be needed to address this question further.

Finally, the ANOVA revealed that the predictions made by the floor-vibration monitoring system also showed less deviation than those of the 2 research-grade ActiGraph GT9X monitors for both the number of steps and activity intensity end points (Figure 5). The underestimation of the number of steps noted for the waist-worn ActiGraph device may be related to walking gait characteristics when movements are performed in closed and narrow spaces. Such environments may not allow sufficient acceleration to meet the necessary signal processing threshold criteria required to count a step, as suggested elsewhere [32]. In contrast, the overestimated number of steps observed for the wrist-worn device may be the result of confounding upper limb movements performed in a frequency range similar to that of walking gait, which may occur in the course of completing housework-related activities.

Regardless of the performance of the ActiGraph GT9X monitors and although no external validation experiment has been conducted, taken together, these observations emphasize the good performance of the floor monitoring system for the quantitative evaluation of physical behaviors performed in home settings.

Although the market for wearable activity trackers is still in its growing phase [6], waist- and wrist-worn physical activity monitors, including research- or consumer-grade devices, have been associated with inaccurate predictions of daily physical activity [10-12]. During the past decade, the computation of accurate predictions for housework-related activities using traditional accelerometer-based activity tracker devices has been the object of specific software development [33,34]. However, this study still showed statistical differences between the predictions made by the ActiGraph GT9X monitors and the actual values for both number of steps and activity intensities. These observations emphasize the necessity of developing new methods that can accurately evaluate physical behaviors at home to improve the computation of daily physical activity metrics. When considering long-term use, it is crucial to distinguish between consumer- and research-grade devices. This study used 2 ActiGraph GT9X monitors, recognized as research-grade devices, in the context of a short semistructured experiment. However, consumer-grade physical activity tracking devices used in everyday life are subject to additional extrinsic limitations that can impede their ability to provide continuous monitoring. For instance, the common practice of removing watches and other wearables at home can have significant impacts on the evaluation of physical behaviors in home settings.

Considering the current limitations of wearable activity tracker devices, smart home systems, such as floor-vibration monitoring technologies similar to those used in this study, present a suitable opportunity to enhance the self-monitoring of daily physical activity. Such systems offer a novel approach to improve the accuracy of estimating energy expenditure and the number of steps performed at home, especially when considering their integration with a 5G network composed of interconnected devices dedicated to the evaluation of daily physical activity. By ensuring accurate and continuous measurements when individuals are at home, smart home systems could help maintain people’s interest in self-monitoring, making them a pivotal factor in promoting and sustaining active and healthy lifestyles. However, the potential widespread adoption of such systems should not only be considered from a technological perspective but should also acknowledge the role of sociocultural factors in shaping user acceptance and usability.

The main limitation of this study is that the proposed method only assesses the physical behaviors of 1 inhabitant at a time. Quantifying the physical behaviors of multiple individuals would require additional signal processing tools to link vibration events with the individuals generating them. Although each individual may exhibit a unique gait signature reflected in the floor-vibration signal, extracting such information is beyond the scope of this study. This limitation could also be addressed by analyzing the sequence of interactions with smart and connected home furniture devices, similar to what has already been described elsewhere [35]. Furthermore, it is important to note that the experimental Ocha-House used in this study was originally designed for a single inhabitant, aligning with the living environment of millions of Japanese people. The structured nature of the experimental protocol may be cited as a second limitation of the study, which restricts the generalization of the observations to what may happen under free-living conditions. To further evaluate the feasibility of using the floor-vibration–based monitoring method, semistructured experiments using a portable breath-by-breath gas exchange analyzer could be conducted to assess energy expenditure during longer periods of activity. A third limitation is that the external validity of the floor-vibration parameter–based activity intensity prediction models (Table 2) was not tested, thus mitigating the interpretations of the comparison test performed against the actigraphy method. The cost of the system is also considered to be a limitation. In this study, expensive, high-sensitivity shear-type accelerometer sensors were used. Further studies are necessary to explore the feasibility of using cheaper accelerometer sensors similar to those commonly used in wearable devices. The results of the multiple regression analyses (Table 2; Multimedia Appendix 2) indicated that the floor-vibration–based step count and moving distance parameters contributed more to the activity intensity prediction models. These 2 parameters may not require the computation of a high-resolution signal. Fourth, the quasi-absence of responses in the floor count, step count, and moving distance parameters during sitting and watching videos may suggest that the floor-vibration monitoring system may also be capable of evaluating sitting behaviors (Figures 4E-Figures 4G). However, owing to the structured nature of the protocol, which involves short observation windows, further interpretations regarding the accuracy of the system in predicting energy expenditure for home-based sitting activities cannot be made. Given the importance, complexity, and intricacies of sedentary behaviors that can occur at home, specific studies should be designed to understand how floor-vibration monitoring systems may contribute to the objective assessment of home-based sedentary behaviors. Finally, the participants in this study were all women and exhibited relatively homogeneous characteristics in terms of age and weight. The BMI scores indicated a limited variability in physical fitness. These observations constrain the generalizability of our results to a more diverse population. Given that age and physical fitness are recognized determinants of energy expenditure, future investigations should aim to recruit a more diverse sample of participants and consider a broader spectrum of personal characteristics in the development of energy expenditure prediction models.

This study has several strengths and originalities. First, the results of the present experiment are in line with those of previous studies, which described a good relationship between the force exerted by an individual on the floor of a small squared 6.25 m² metabolic chamber equipped with a force transducer and the actual energy expenditure [36-38]. They allow extending the previous observation to a different sensing technology, larger non–squared living surfaces, and a wider range of activities that are usually performed at home. Another strength of this study is that it is the first to compare the outcomes of a smart home system with those of research-grade activity trackers.

This study presents a novel floor-vibration monitoring system that can be used in smart home settings to quantify physical activity at home. The hardware included high-sensitivity accelerometers. In this case, 8 sensors were required to cover a surface area of 42 m². The software includes a simple data processing workflow for the computation of the floor count parameter, which is a quantitative index of the floor vibrations, and the step count and moving distance parameters. Regression models combining the information of these 3 parameters showed a strong association with the actual intensities measured using indirect calorimetry for the 4 tested home-based activities. A significant association was also observed between the step count parameter computed using the floor-vibration signal and the actual number of steps. Further studies, conducted under real-life conditions or using semistructured experimental protocols, are necessary to extend the results of this study and validate the monitoring of floor vibrations as a surrogate method for evaluating physical behaviors at home. Considering the current evolution of 5G technologies and IoT devices, smart home systems are expected to contribute to a more continuous evaluation of daily physical activity.