Study of Human Outdoor Walking with a Low-Cost GPS and Simple Spreadsheet Analysis : Medicine & Science in Sports & Exercise

Journal Logo

BASIC SCIENCES: Symposium: Exercise, Antioxidants, and Cardioprotection

Study of Human Outdoor Walking with a Low-Cost GPS and Simple Spreadsheet Analysis


Author Information
Medicine & Science in Sports & Exercise 39(9):p 1570-1578, September 2007. | DOI: 10.1249/mss.0b013e3180cc20c7
  • Free


The effect of physical activity on public health is well recognized (5,14,15). Among a broad range of free-living human activities, walking practice has been reported to be a key activity to achieve the benefits of physical activity on health outcomes (3,5,9,11). Self-reported free-living walking practice mainly relies on outdoor walking (9,11) and represents a type of behavior somewhat difficult to characterize because of differences in intensity, duration, frequency, and the related volume of energy expenditure. Most of the time, the assessment of free-living outdoor walking has been performed by self-reported questionnaires, which inevitably leads to misclassification (11,19). That is why objective tools should be evaluated, to assess with accuracy the pattern of free-living outdoor walking and to reliably analyze its relationships with specific clinical outcomes (1).

For this purpose, high-sampling rate accelerometers are reliable tools for accurately detecting daily physical activities such as walking (12,13), allowing one to identify both walking and resting periods. Nevertheless, various limitations have been reported concerning the prediction of speed when using accelerometers, including the large intersubject variability of the relationship between speed and raw acceleration or accelerometer counts (16,19), the complexity of the algorithms used to convert raw acceleration (16,19,23), and the inability to predict speed in case of slope changes (16,23) or to measure pathological gaits such as claudication (23).

The global positioning system (GPS) is used in biomechanical and sports physiology studies because of its precise determination of position overground (7,8,17,18,21-24,26,27). The available studies that have tested the accuracy of the GPS to assess walking have focused on the determination of walking speed overground (17,18,22), regardless of walking and resting bouts. This is probably not sufficient if one aims at studying precisely the pattern of outdoor walking and eventual limitations to walking capacity. Contrary to the accelerometers, GPS seems to have various potential advantages for studying outdoor walking under free-living conditions, by allowing direct position estimation, measurements of ground slopes, and calculation of speed without individual calibration before use (23). Heretofore, differential and high-sampling rate GPS have been used in research reports studying human locomotion, but these highly precise devices, although they provide a highly accurate position determination (27), cannot be used routinely because of their cost and physical size. In contrast, low-cost, relatively small GPS devices are commercially available. The accuracy of such devices in studying patterns of outdoor walking should be assessed.

The main objective of this study was to test whether the Garmin 60 GPS device was accurate enough to assess outdoor walking during prescribed walking protocols. For this purpose, analysis of data downloaded from the GPS should accurately detect walking and resting periods and accurately estimate the speed of the corresponding walking periods. To achieve this objective, an emphasis was put on the study and comparison of different signal-processing techniques (low- and high-pass filters) to correct original signals downloaded from the GPS receiver. Our study aims at determining the validity of this low-cost, commercially available GPS for studying outdoor walking, which is a useful step towards using the GPS technology for further studies of free-living walking practice.



A single, low-cost (approximately $470) GPS receiver/logger (Garmin GPS 60), which includes the European Geostationary Navigation Overlay Service (EGNOS) function, was used to determine the positions of the subjects during all our experiments. The principles of GPS and EGNOS-enabled GPS have been well described elsewhere (19,24,27). In brief, the GPS is a satellite-based navigation system made up of a network of 24 operational satellites in orbit around the earth. Each satellite broadcasts radio frequency signals that are decoded by the receiver. The GPS receiver calculates the difference between the moment when a signal was transmitted by a satellite and the moment when it was received and, by multiplying travel time by the speed of light, the exact distance to each satellite is calculated. An exact, three-dimensional position is obtained by trigonometry, through the calculation of the distance to four satellites. The EGNOS system improves position determination (± 1-3 m) by collecting error-correction data from multiple reference stations via terrestrial communications. The error-correction data are transmitted to an additional Geostationary Earth Orbit Satellite, after which they are retransmitted to the GPS receiver. The advantage is that the error-correction data can be received via the same antenna as the standard GPS signal, without increasing the weight of the device. The EGNOS-enabled GPS receiver has been shown to improve the determination of position and speed overground compared with non-EGNOS units (27).

The GPS receiver (15.5-cm height × 6.10-cm width × 3.30-cm depth; weight: 198 g with two batteries) was placed in the external pocket of a backpack. An external receiving antenna was placed over the backpack to improve the signal reception. Because our future goal was to perform prolonged recordings (8-10 h), considering the memory capacity of the device, the chosen recording rate was 0.5 Hz. Throughout all experiments, subjects were asked to wait for approximately 10 min, to allow for GPS initialization (time for the initial detection of the satellites), before the experiments could be started. Recorded positions (latitude, longitude, altitude) were downloaded from the GPS receiver after each experiment using the MapSource software (version 6, Garmin Ltd.), and automatically converted by this software to speeds and distances. The speed values obtained were analyzed on a personal computer using a spreadsheet (Microsoft Excel 2000). Throughout this article, results are expressed as mean ± SD.

Study Design

Our approach attempts to demonstrate the relevance of the GPS with our specific sampling configuration, in the estimation of outdoor walking. For this purpose, the present study was divided into three consecutive experiments.

  • In experiment 1, the accuracy of various signal processing was determined, to explore the ability of the GPS technique to detect speed in a chain of walking and resting bouts of different durations from a prescribed walking protocol (PWP).
  • Experiment 2 was designed to determine the accuracy of arbitrarily defined high-pass filters as compared with the methodologies derived from experiment 1, through a blinded analysis of other PWP. This experiment was required to confirm the validity and applicability of the methodologies developed in experiment 1 in a new series of PWP.
  • Experiment 3 was the application for the prediction of walking speed and distance of the best methodology developed. Speed and distance derived from the analysis of GPS recordings during a new series of walking bouts of various distances were compared with actual speed (directly measured by chronometry) and distance.

This study was approved by our institutional ethics committee. Subjects were informed of the experimental procedure, and informed consent was obtained from all participants.

Experiment 1: Determination of Processing Methodologies


Six healthy subjects (32 ± 14 yr, 173 ± 8 cm, 64 ± 11 kg) equipped with the GPS receiver were asked to perform a PWP of 31 min 30 s, three or four times, on an outdoor athletic track (400 m long, flat area, free of compact trees, free of buildings, with just a gallery that runs parallel to one side of the track). Each PWP arbitrarily included a succession of six prescribed walking bouts of 2 min, 4 min, 30 s, 15 s, 1 min, and 8 min, separated with resting bouts of 30 s, 15 s, 4 min, 2 min, 1 min, and 8 min. Subjects walked at a freely chosen speed on the interior lane of the athletic track and were closely followed and directed throughout the procedure by an investigator. Onset of each walking and resting bout was determined with a stopwatch and a verbal call such as "go" and "stop" for walking and resting, respectively. The GPS recording started and ended about 15 min before and after the PWP. The period of interest was detected with both recorded time and through the simultaneous validation of an event marker on the GPS, before the GPS was placed in the backpack.

Data processing.

Many mathematical methods have been proposed in the literature to reduce the signal-to-noise ratio (4,20,25). Our attempt was to propose a processing methodology that requires no specific knowledge in signal-processing methods and that can be used easily in field application, using a basic spreadsheet. The initial stage of this analysis was to calculate the mean and standard deviation (SD) of the individual walking speed (IWS). Mean and SD of IWS were arbitrarily calculated during the first walking bout (2 min) of the protocol and were used as a basis to correct GPS signal instabilities in walking and resting bouts of the PWT.

Indeed, during preliminary recordings, it was observed that

  1. An artifact of high value at the beginning of walking bouts was often generated (i.e., the first speed value after a resting bout).
  2. Although close to it, raw speed data were seldom equal to zero during resting bouts. This confirms previous studies showing that GPS was unable to detect static activity (17,21).
  3. Isolated, high, false speed values may occur at random during PWT, and very short stops may occur because of external events interfering with the subject's walk.

Typical examples of these signal instabilities are illustrated in Figure 1, showing two typical recordings of the 21 available ones. The preliminary observations of signal instabilities have been used as a basis for the different processing steps developed in the present experiment. Then, to remove the artifact of high value at the beginning of walking bouts, a low-pass filter (LPF) designed to transmit a given signal below a certain value, while excluding the signal above this fixed value, was defined. It was arbitrarily fixed at two times the IWS of the first 2 min. Above this limit, raw datum was replaced by the average of the following five raw data. Then, because raw speed data were seldom equal to zero, a high-pass filter (HPF), designed to transmit a given signal above a certain value, while excluding the signal below this fixed value, had to be defined. Generally, HPF are fixed values. The choice was made to define an HPF that was not a fixed value, because the walking speed can be different among subjects, particularly in older subjects and/or in disabled subjects (with intermittent claudication). However, a low value will generate many false "go" and few false "stops," and vice versa. Thus, to define the best compromise between false "go" and false "stop" detection, and to account for individual walking speed, five different values for the HPF, corresponding to the calculated IWS minus 2, 3, 4, 5, or 6 times its standard deviation, were tested. Below each tested HPF, raw datum was shifted to zero. Finally, to remove the isolated high false speed values and eventual very short stops, an artifact processing (AP) was defined, to convert all resting and walking bouts that lasted 2 or 4 s, because such small bouts were expected not to be detected with a 0.5-Hz sampling rate and to be of little significance. Thus, "stopping bouts" ≤ 2 s were converted to "walking bouts" by shifting the raw datum N (= 0) to the mean of N − 2, N − 1, N + 2, N + 3. The procedure was repeated twice (to eliminate bouts of 4 s; i.e., two raw data). Consistently, "walking bouts" ≤ 4 s were converted to "stopping bouts" by shifting the raw data to zero.

Example of raw speed data provided by the MapSource software from downloaded GPS recordings (upper and lower panel) obtained from the same prescribed protocol in two different subjects during experiment 1. The figure shows typical signal instabilities (see the Methods section for a precise description of these signal instabilities).

Experiment 2: Validation Experiment


Ten subjects (23 ± 2 yr, 173 ± 8 cm, 65 ± 10 kg) were equipped with a watch, the GPS receiver (backpack), and an MP3 player, over which a PWP (different for each subject) was vocally recorded, including both prescribed walking and resting bouts. The recording was started at the hospital, with recording of time and validation of an event marker, as for experiment 1. Thereafter, they were instructed to go to a designated public park near the hospital to perform the PWP. This public park is a flat area (no hills) and is free of motorized vehicle circulation, buildings, and compact trees. Once the subjects were in the park and ready to walk and after the 10 min of GPS initialization, they turned on the MP3 player, noted the exact time of the beginning of the protocol, and then carefully performed the PWP at a freely chosen walking speed. Each protocol lasted between 18 and 20 min. The recording was stopped when the subject came back to the hospital.

The rationale for the determination of the number of bouts for each protocol was as follows. After experiment 1, the highest accuracy for the bout-level analysis from GPS signal was 92.9% (see Results section below). Assuming that the accuracy resulting from the manual postprocessing (see Data analysis section, below) would be approximately 99%, using a binomial distribution (10), the minimal number of bouts to be performed in experiment 2 to reach a one-sided statistical significance level of 0.05 and 80% power was 160. The choice was made to prescribe 10 different PWP (10 subjects) including, in a random order, from 8 to 10 walking and from 8 to 10 resting bouts each.

Data processing.

Various methodologies were used, from sole fixed HPF to multistep approaches as defined in experiment 1. The aim was to confirm that methods derived from experiment 1 were applicable and provided better results than fixed cutoff values. Additionally, the accuracy of manual postprocessing, associated with one of the methodologies developed in experiment 1, was assessed.

Processing methodologies 1 and 2.

The aim was to evaluate the accuracy of two different fixed HPF, arbitrarily defined at 0.5 and 1.0 km·h−1, respectively. Below these cutoffs, raw speed data were converted to zero. No LPF or artifact processing was used.

Processing methodology 3.

This methodology consisted of applying the procedure described in experiment 1, using a fixed HPF of IWS minus 3.5 times the SD of IWS. The rationale for this fixed HPF is that, on average, it provided the best compromise of walking bouts and resting bouts incorrectly determined (see Results of experiment 1).

Processing methodology 4.

This methodology is only slightly different from the methodology 3. Here, the HPF was adapted from an algorithm that accounts for the variability of the GPS signal. The latter was calculated as a coefficient of variation from IWS and SD of the two first minutes of the protocol (CVIWS = SD/IWS × 100). Indeed, if CVIWS is high, there might be a risk that IWS minus 3.5 SD would be below zero. The rationale for this algorithm is explained in the Results section of experiment 1.

Processing methodology 5.

After methodology 4 processing was completed, two investigators were asked to perform a manual graphic comparison between processed signal and raw signal (manual postprocessing). They were only informed that no resting or walking bout of less than 5 s was prescribed in any of the walking protocols. They were only allowed to 1) remove any persisting artifact data for bouts ≤ 4 s; and 2) check that the first walking bout of 2 min used for the calculation of IWS reflected the mean speed of the overall detected walking protocol; if not, slower walking bout could be misdetected, requiring a change and adaptation of the HPF.

Statistical Data Analysis of Experiments 1 and 2

To analyze the accuracy of each tested parameter of the processing methodology, two types of analyses were performed: an analysis at the "sample level" and an analysis at the "bout level." These analyses relied on the comparison between the PWP, considered as the gold standard, and the walking protocol detected by the GPS and processed following the different methodologies of signal processing.

First, the sample-level analysis relied on the comparison between the actual and detected protocol for each datum (i.e., sample by sample). This procedure enabled one to calculate 1) "true" walking samples (with correspondence between actual and detected walking samples); 2) "true" resting samples; 3) "false" walking samples (actual resting samples detected as walking samples); and 4) "false" resting samples. Thus, the number of both walking and resting samples that was accurately detected could be calculated. Using this procedure, for some walking trials, the HPF could be so low that resting bouts could not be detected, resulting in a high percentage of false walking samples. Consequently, for those walking trials, the calculation of false bouts would be impossible, because a few resting bouts could not be detected. As a result, a walking trial was excluded from the calculation of false bouts when the accuracy at the sample level was < to 50% (because of a high percentage of false walking samples). Note that for the sample-level analysis, both walking and resting samples were renamed "go" samples and "stop" samples, respectively, to avoid any confusion with the bout-level analysis.

Secondly, the bout-level analysis relied on the comparison of the number of bouts between the prescribed and detected walking protocols. Indeed, the same equal number of sample detection errors may result in different bout-detection ability, depending on how the sample errors are distributed through the recording. The rationale for the bout-level analysis was to calculate the number of both walking and resting bouts that were detected and to compare it with the PWP. Using this procedure, the first step was to calculate the absolute number of both false resting and false walking bouts, which could be greater than the number of walking or resting bouts actually performed (i.e., multiple false walks may occur during a single resting period). The final end point was to calculate the accuracy of the GPS processing methodologies in the detection of walking and resting bouts actually performed. For this purpose, the number of incorrectly detected bouts was calculated (i.e., two false rests in a given walking bout only result in a single walking bout incorrectly detected). Note that, contrary to the sample level, the calculation of the accuracy here was possible with no exclusion of walking trials. Indeed, in this validation study, for the subjects with a high percentage of "false go" and a low accuracy at the sample level, the number of resting bouts incorrectly detected was automatically equal to the number of resting bouts included into a given protocol (e.g., six resting bouts for protocol of experiment 1), because no resting bouts was detected. Accuracies for each tested methodology, obtained after bout-level analysis, are reported with 95% confidence intervals. In experiment 2, comparison of the accuracies between the different tested methodologies was analyzed using the MacNemar test comparing true and false classified bouts with the different processing (SPSS statistical software, version 12.0.1, 2003).

Experiment 3: Application for the Prediction of Walking Speed and Distance


Fourteen subjects (22 ± 2 yr, 176 ± 5 cm, 70 ± 6 kg) performed twice a PWP of 2000 m long, including two series of 100, 200, 300, and 400 m, respectively, in a random order (different for each subject). PWP were performed on the same outdoor athletic track as in experiment 1. Four blocks were placed every 100 m. Subjects were asked to be careful to walk on the interior lane of the athletic track at a freely chosen speed, and to stop at the next block when hearing a whistle blown by the investigator 10 m before the block. In this way, prescribed resting bouts of 30 s were inserted between each walking bout (series). For nine subjects (22 ± 3 yr, 177 ± 6 cm, 69 ± 6 kg), the time required to perform each walking bout was assessed by chronometry (Geonaute Trt'L 500, Decathlon Ltd., France) and was carefully recorded by the investigator, allowing the calculation of the actual speed. For each PWP, the recording was started on the athletic track with recording of time and validation of an event marker.

Data processing.

Raw speed and distance calculated by the MapSource software from the GPS data were processed according to processing methodology 5 (see above). Thereafter, processed speed and distance data were compared with actual speed (measured by chronometry) and distance.

Statistical Data Analysis of Experiment 3

The comparison between actual speed/distance and speed/distance generated by the processing methodology was performed using regression analysis, as well as graphically with Bland-Altman plots (2) using the ratio (not the difference) between the actual distance or speed and calculated distance or speed against their mean (GraphPad Prism software, version 4.01, 2004, GraphPad Software Inc.). The typical error of the measurement (TEM) and the coefficient of variation (CV) within walking bouts were also calculated according to the statistical procedure proposed by Hopkins (6).


Experiment 1: determination of processing methodologies.

The PWP was performed 21 times for a total of 126 walking bouts and 126 resting bouts. The average walking speed for the first 2 min of the 21 detected walking protocols was 4.9 ± 0.6 km·h−1, ranging from 3.5 to 6.3 km·h−1. As previously shown, Figure 1 represents a typical result of the raw speed data downloaded from the GPS receiver and provided by the MapSource software during two walking protocols.

At the sample level, it was of interest to note that

  1. The accuracy after the LPF alone was 50.5%.
  2. Adding an HPF, accuracies ranged from 83.1% (IWS minus 6 SD) to 97.2% (IWS minus 3 SD).
  3. Then, the addition of the artifact processing resulted in an accuracy ranging from 83.3 to 98.2% (Table 1).
Results of the sample-level analysis according to different methodologies of signal processing of GPS speed data after experiment 1.

Accuracy decreases as the SD of the HPF increases, but only when CVIWS is above 15% (data not shown). To be more specific, with CVIWS below 15%, higher accuracy is obtained using an HPF that was fixed to IWS minus 5 SD for IWS. Otherwise, when CVIWS was ≥ 15%, higher accuracies were obtained using an HPF that was fixed to IWS minus 2 SD for IWS. These observations justify the use of an adapted HPF (algorithm) for experiment 2, as previously indicated in the Methods section.

At the bout level, the number of absolute false bouts ranged from 31 to 62 (Table 2), recalling that any single of the 252 bouts could be divided into many ones after signal processing. This clearly shows that the calculation of the accuracy at the sample level was not sufficient to determine the validity of the GPS for detecting walking and resting bouts, because false bouts could be generated despite a high calculated accuracy at the sample level. One should keep in mind that a single incorrectly detected bout may include many false events, as reported in Table 2. Then, calculating the number of incorrectly detected bouts, including all the detected walking protocols, the accuracy ranged from 79.4 to 92.9% (Table 3). The optimal theoretical value to be applied to the SD of IWS for the HPF, to obtain the lowest number of incorrectly determined bouts, was expected to be between 3 and 4; 3.5 was chosen arbitrarily. This was obtained by plotting the number of incorrectly determined walking bouts versus the number of incorrectly determined resting bouts from the results of Table 3 (not shown). This value has been used as a rationale for the application of a fixed HPF of 3.5 SD for experiment 2 (see Methods section).

Number of false bouts according to different methodologies of signal processing of GPS speed data after the bout-level analysis of experiment 1.
Number of incorrectly detected bouts and corresponding accuracies according to different methodologies of signal processing of GPS speed data after the bout-level analysis of experiment 1.

Experiment 2: validation experiment.

The total number of bouts performed in experiment 2 was 176 (91 walking bouts and 85 resting bouts). Results of experiment 2 are shown in Tables 4 and 5. Accuracies achieved with the simple cutoffs were statistically lower than with the artifact processing/adapted HPF methodology (P < 0.01). As expected, the best signal processing was achieved through the manual postprocessing, providing 1.1% of walking bouts incorrectly determined and 4.7% of resting bouts incorrectly determined. All false bouts that were not shifted by the investigators lasted less than 15 s. The accuracy obtained with the manual postprocessing was significantly higher than that with other methodologies (P < 0.05).

Number of false bouts according to different methodologies of signal processing of GPS speed data after the bout-level analysis of experiment 2.
Number of incorrectly detected bouts and corresponding accuracies according to different methodologies of signal processing of GPS speed data after the bout-level analysis of experiment 2.

Experiment 3: application for the prediction of walking speed and distance.

A total of 252 bouts were performed, among which 162 were timed. One bout (100 m) was excluded from the analysis by the investigator because of interference in the GPS signal. Table 6 displays the typical error of the measurement (TEM) and the coefficient of variation (CV) for distance and speed. As shown, the CV for the processed distance and speed prediction decreased when the covered distance increased. Figure 2 shows that there was an excellent relationship between actual and processed distance or speed. From the Bland-Altman plots (Fig. 3), it should be noted that the relative errors of distance prediction decreased as the covered distance increased, whereas no relationship was found between the relative error in speed determination and the average speed.

Accuracy of the estimation of distance and speed processed from downloaded GPS data according to the covered distance; the typical error of the measurement (TEM) and the coefficient of variation (CV) are presented.
Relationships between actual vs processed distance (A) and speed (B) in experiment 3. Regression analyses and coefficients are presented.
Bland-Altman plots of the relative errors of actual vs processed distance (A) and speed (B) in experiment 3. Full lines represent the bias, and dotted lines represent the 95% limits of agreement (bias ± 1.96 SD).


Detection of walking and resting bouts.

Because the study of walking patterns first requires detecting the corresponding walking periods, the first step in this study was to focus on the detection of walking and resting bouts. On the basis of the results of experiments 1 and 2, there were four major findings concerning this issue: 1) GPS raw speed data cannot, per se, enable one to detect walking and resting bouts, because of signal instability; 2) simple cutoffs at a fixed value are not sufficiently accurate to detect walking and resting periods; 3) the combination of an LPF, an adapted HPF, and artifact processing is required to accurately detect walking and resting bouts; and 4) a manual postprocessing to complete previous automatic processing methodology provides the highest concordance with the real values. To the best of our knowledge, this is the first study to develop a relatively simple signal processing using a spreadsheet, designed specifically to detect walking and resting bouts using a commercially available, low-cost GPS. It could be argued that five bouts were incorrectly detected after the manual postprocessing. However, most bouts incorrectly detected by the investigators were bouts of less than 15 s. Whether such short bouts account for a substantial part of the total daily walking activity should be the object of further investigation.

To date, most sports physiology studies using non-high-precision professional GPS receivers have focused on speed of running (7,8), with no mention of a specific processing method to detect running bouts. Nevertheless, the detection of walking bouts (and resting bouts) over a given period was expected to be more complex than the detection of running bouts. Indeed, the speed difference between walking and resting conditions is lower compared with that for running conditions, and the GPS technique fails to measure static activities (17).

The available studies that tested the accuracy of professional GPS in assessing walking have focused on the determination of walking speed overground (17,18,22), but they rarely have analyzed the detection of walking and resting bouts directly (21). Further, in this latter study (21), the authors used a differential GPS and focused on the analysis of the gait pattern during walking. A previous report (19) suggests that people "spontaneously select a comfortable pace" between 4 and 6 km·h−1. This contrasts with the broad range of reported walking paces in long-term observational studies (9,11). It is clear that healthy elderly subjects or subjects with walking disabilities may exhibit slower speeds. GPS recording requires no calibration to measure individual speed. Further, the signal processing developed in the present study is based on an individual's walking speed; thus, it is expected to account for the different populations to be recorded. This would require future confirmation.

GPS prediction of walking speed and distance.

Using a nondifferential GPS, the speed-prediction error for walking (SD of the error) was previously reported as high as 1.1 km·h−1 (17), which is substantially greater than our results. The TEM and CV we found for walking bouts of 100 m were 0.22 and 3.40% respectively. This is in the same range as those reported using differential GPS (18) in a very small number of walking bouts (N = 12), with one subject, on a single day (SD ranging from 0.08 to 0.15 and CV ranging from 1.38 to 2.67% for speed walking prediction, according to the methods of speed calculation). Another interesting finding is that the speed and distance relative error decreased (CV) when the covered distance increased (Table 6). For free-living monitoring, this is useful information because there is a substantial number of types of gait, with continuous and discontinuous walking, resulting in a broad range of covered distance.

Study limitations.

In the present report, the accuracy of the GPS in detecting walking and resting periods was tested from PWP. As a consequence, one could argue that there are clear limits on what can be achieved while monitoring completely free-living walking during a prolonged period. Outdoor activities are relatively easy to characterize. The analysis of speed is expected to be sufficient to differentiate walking from other activities such as running or motorized transports. Additionally, the use of altitude measurements could be useful to discriminate other activities such as walking up or down (e.g., staircases or slope changes).

Our study was performed in areas with no major environmental interferences (trees, buildings, hills). It cannot be excluded that experiments performed in different environmental conditions or geographic areas would modify the results.

It could be suggested that different atmospheric conditions may influence satellite reception. To the best of our knowledge, weather conditions (wind, clouds, rain) show very little influence on the quality of GPS reception (23), whereas it is true that differences in the number of available satellites (e.g., from one day to another, as in the present study) may influence speed error prediction (27). This is inherent to outdoor measurements.

The fact that the GPS data downloaded from the receiver were reinterpreted by the software program used (MapSource) must also be discussed. Indeed, when downloaded, the three-dimensional GPS data (latitude, longitude, altitude) stored by the receiver are translated to linear measures of speed and distance, using an unknown specific algorithm. Software companies consider their algorithms proprietary, so it is unclear whether our results could be influenced by the algorithm used by the MapSource software. It is possible that our results are only valid with the specific receiver and software used, and that they cannot be extended to other devices.

It is likely that the chosen 0.5-Hz sample frequency for storing GPS data in the receiver was responsible for a low signal-to-noise ratio in our data, at least in some subjects, with lower walking speed. For example, the slowest self-selected walking speed was 3.5 km·h−1, which equals 1.94 m every 2 s. This value is easily within the ± 1-3 m of true GPS measurement error. This could be a contributor to the "signal instability" observed, and this also matches well with the observation that the speed and distance-prediction error decreased when the covered distance increased (i.e., when the signal began to greatly exceed the noise). Nevertheless, this was the highest sampling frequency consistent with the device memory capacity and with our final goal, which was to perform prolonged (8-10 h) outdoor measurements in normal and diseased subjects.


A low-cost, commercially available GPS is accurate in studying outdoor walking, allowing one to identify walking and resting periods and to accurately estimate both the speed and distance of the corresponding walking bouts for healthy subjects. The GPS technology seems to be a promising tool for the objective measurement of outdoor walking, provided that the recording may cover a prolonged period of time. It has potential medical applications for the analysis of walking activity and walking capacity. The ability to record walking activity would be of great interest in studying the relationship between objectively measured outdoor walking and clinical outcomes (1). Estimation of free-living walking capacity is also an interesting issue in patients with walking disability, such as intermittent claudication. Indeed, walking capacity of vascular patients with claudication under free-living conditions remains mostly undetermined.

The authors are indebted to Nicolas Quirion and F. Debonnaire for their help in data analysis, to N. Boileau who corrected the grammar and style of the manuscript, and to Pr L. Leger for his helpful suggestions and comments.

Alexis Le Faucheur is a recipient of a grant from the Conseil Général du Maine-et-Loire. The present study was granted in part by the Société Française de Médecine Vasculaire and was promoted by the University Hospital in Angers.


1. Blair, S. N., and W. L. Haskell. Objectively measured physical activity and mortality in older adults. JAMA 296:216-218, 2006.
2. Bland, J. M., and D. G. Altman. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1:307-310, 1986.
3. Duncan, G. E., S. D. Anton, S. J. Sydeman, et al. Prescribing exercise at varied levels of intensity and frequency: a randomized trial. Arch. Intern. Med. 165:2362-2369, 2005.
4. Fraden, J. Destruction of noise in biotelemetry. Biotelem. Patient Monit. 6:207-221, 1979.
5. Hakim, A. A., H. Petrovitch, C. M. Burchfiel, et al. Effects of walking on mortality among nonsmoking retired men. N. Engl. J. Med. 338:94-99, 1998.
6. Hopkins, W. G. Measures of reliability in sports medicine and science. Sports Med. 30:1-15, 2000.
7. Larsson, P. Global positioning system and sport-specific testing. Sports Med. 33:1093-1101, 2003.
8. Larsson, P., and K. Henriksson-Larsen. The use of dGPS and simultaneous metabolic measurements during orienteering. Med. Sci. Sports Exerc. 33:1919-1924, 2001.
9. Lee, I. M., K. M. Rexrode, N. R. Cook, J. E. Manson, and J. E. Buring. Physical activity and coronary heart disease in women: is "no pain, no gain" passe? JAMA 285:1447-1454, 2001.
10. Lenth, R. V. Some practical guidelines for effective sample-size determination. Am. Stat. 55:187-193, 2001.
11. Manson, J. E., P. Greenland, A. Z. LaCroix, et al. Walking compared with vigorous exercise for the prevention of cardiovascular events in women. N. Engl. J. Med. 347:716-725, 2002.
12. Mathie, M. J., B. G. Celler, N. H. Lovell, and A. C. Coster. Classification of basic daily movements using a triaxial accelerometer. Med. Biol. Eng. Comput. 42:679-687, 2004.
13. Mathie, M. J., A. C. Coster, N. H. Lovell, and B. G. Celler. Detection of daily physical activities using a triaxial accelerometer. Med. Biol. Eng. Comput. 41:296-301, 2003.
14. Paffenbarger, R. S., Jr., R. T. Hyde, A. L. Wing, I. M. Lee, D. L. Jung, and J. B. Kampert. The association of changes in physical-activity level and other lifestyle characteristics with mortality among men. N. Engl. J. Med. 328:538-545, 1993.
15. Pate, R. R., M. Pratt, S. N. Blair, et al. Physical activity and public health. A recommendation from the Centers for Disease Control and Prevention and the American College of Sports Medicine. JAMA 273:402-407, 1995.
16. Perrin, O., P. Terrier, Q. Ladetto, B. Merminod, and Y. Schutz. Improvement of walking speed prediction by accelerometry and altimetry, validated by satellite positioning. Med. Biol. Eng. Comput. 38:164-168, 2000.
17. Schutz, Y., and A. Chambaz. Could a satellite-based navigation system (GPS) be used to assess the physical activity of individuals on earth? Eur. J. Clin. Nutr. 51:338-339, 1997.
18. Schutz, Y., and R. Herren. Assessment of speed of human locomotion using a differential satellite global positioning system. Med. Sci. Sports Exerc. 32:642-646, 2000.
19. Schutz, Y., S. Weinsier, P. Terrier, and D. Durrer. A new accelerometric method to assess the daily walking practice. Int. J. Obes. Relat. Metab. Disord. 26:111-118, 2002.
20. Taulu, S., M. Kajola, and J. Simola. Suppression of interference and artifacts by the signal space separation method. Brain Topogr. 16:269-275, 2004.
21. Terrier, P., Q. Ladetto, B. Merminod, and Y. Schutz. High-precision satellite positioning system as a new tool to study the biomechanics of human locomotion. J. Biomech. 33:1717-1722, 2000.
22. Terrier, P., Q. Ladetto, B. Merminod, and Y. Schutz. Measurement of the mechanical power of walking by satellite positioning system (GPS). Med. Sci. Sports Exerc. 33:1912-1918, 2001.
23. Terrier, P., and Y. Schutz. How useful is satellite positioning system (GPS) to track gait parameters? A review. J. Neuroengineering Rehabil. 2:28, 2005.
24. Terrier, P., and Y. Schutz. Variability of gait patterns during unconstrained walking assessed by satellite positioning (GPS). Eur. J. Appl. Physiol. 90:554-561, 2003.
25. Veltink, P. H., H. B. Bussmann, W. de Vries, W. L. Martens, and R. C. Van Lummel. Detection of static and dynamic activities using uniaxial accelerometers. IEEE Trans. Rehabil. Eng. 4:375-385, 1996.
26. Witte, T. H., and A. M. Wilson. Accuracy of non-differential GPS for the determination of speed over ground. J. Biomech. 37:1891-1898, 2004.
27. Witte, T. H., and A. M. Wilson. Accuracy of WAAS-enabled GPS for the determination of position and speed over ground. J. Biomech. 38:1717-1722, 2005.


© 2007 American College of Sports Medicine