Comparison and adaptation of step length and gait speed estimators from single belt worn accelerometer

positioned on lateral side of the body

Taufique Sayeed, Albert Samà, Andreu Català, Joan Cabestany, Technical Research Centre for Dependency Care and Autonomous Living (CETpD)

Rambla de l'Exposicio, 59-69 08800 Vilanova i la Geltru, Spain taufique.sayeed @estudiant.upc.edu, albert.sama@upc.edu, andreu.catala@upc.edu, joan.cabestany@upc.edu

Abstract— Parkinson’s Disease (PD) is a neurodegenerative disease that predominantly alter patients’ motor performance and compromises the speed, the automaticity and fluidity of natural movements. The patients fluctuate between periods in which they can move almost normally for some hours (ON state) and periods with motor disorders (OFF state). Gait properties are affected by the motor state of a patient: reduced stride length, reduced gait speed, increased stride width etc. The ability to assess the motor states (ON/OFF) on a continuous basis for long time without disturbing the patients’ daily life activities is an important component of PD management. An accurate report of motor states could allow clinics to adjust the medication regimen to avoid OFF periods. The real-time monitoring will also allow an online treatment by combining, for instance, with automatic drug-administration pump doses. Many studies have attempted to extract gait properties through a belt-worn single tri-axial accelerometer. In this paper, a user friendly position is proposed to place the accelerometer and six step length estimators are compared considering the proposed sensor placement in a preliminary database of healthy volunteers. Adaptation methods to some of these estimators are also proposed and compared. The comparison shows that the adapted estimators improve the performance and reduce errors in respect of the original methods applied in the new sensor location.

Keywords— Parkinson disease(PD), step length, gait speed, anterior superior iliac spine (ASIS), center of mass (COM),

I. INTRODUCTION

Parkinson’s disease is the second most common neurodegenerative disease in the people aged over 40 years. It is a progressive disorder of the nervous system due to the increasing death of the nerve cells in the substantia nigra located within the basal ganglia of the brain [1]. The death of these specific types of neurons provokes a deficiency of dopamine that alters patients’ motor control, i.e tremor, reduced walking speed, rigidity, muscle stiffness, impaired postural balance etc. Levodopa or similar medication can reduce the motor symptoms of PD patients by increasing the dopamine production. After some years of medication, patients will experience fluctuation between ON and OFF periods. ON periods are characterized as the period when the motor symptoms are almost invisible with the exception of dyskinesia (unvoluntary movements) and the patient feel relatively clear and in control of their movements. OFF state is the period when the motor symptoms are more prominent. As the patients have less control of their movements, they may fall in serious

health risk and sometimes face sudden falls mostly during OFF period. Therefore early detection of OFF states are necessary in order to apply prevention strategies in proper time. Moreover motor status detection can help to improve the response to treatment by modifying the medication regimen according to the motor fluctuations.

Relation between gait parameters and motor states has been widely analyzed in the treatment of PD. Reduced stride length and gait speed are common symptoms on PD patients that will be useful for the diagnose of the motor states [2][3][4][5]. Different expensive clinical tools were used to measure the gait parameters for PD like clinical stride analyzer, Infra-red camera, pressure plates etc. their disadvantage is that, they are not usable outside the laboratory. Recently, some systems using wearable devices with MEMS sensors like accelerometers, gyroscopes and magnetometers have been developed to measure the stride length and gait speed. Their portability and self-administrable methods makes them suitable to be used outside the clinical environment. Using these inertial sensors could help to monitor the patient online and provide proper strategies to handle with the disease[3][5] [6].

Accelerometer data are easy to interpret and less interfering than other sensors, i.e. they barely have drifts due to temperature although drifts appear due to gravity. This is the reason why accelerometers are widely used for continuous gait analysis. Many studies found that positioning the accelerometer on the waist near the joint between the 4th and 5th lumbar vertebrae in the spine (L4-L5) provides then most valuable information for gait analysis than other parts of the body [7][8]. Thus, waist is considered the best position in terms of information to estimate the gait parameters according to the literature. However, Mathie et al [9] showed that this location is impractical to place a device during daily life because it is uncomfortable, may hurt the patient or could be damaged during sitting on a chair or lying on the bed and also needs assistance from other person to find the position and place it in right position. Consequently, they proposed that a different sensor placement above the anterior superior iliac spine (ASIS) could be a better choice. The position is more user friendly and comfortable.

In this work, accelerometer based system is suggested as a potential device placed on the lateral side of waist, the more comfortable place considered to measure the gait parameters to identify the motor status. It is portable, can be used during

daily activities and can also be used to measure the other symptoms of Parkinson’s disease. A review of current methods for step length and gait speed is given in this study. These methods are tested in a small database of signals gathered with the new sensor position on healthy subjects. This paper also presents some adaptation to the current methods for the new position which results are also presented. This work is supported by the Personal Health Device for the Remote and Autonomous Management of Parkinson’s Disease (REMPARK) project[10]. The aim of this project is to develop a Personal Health care System (PHS) for the improved management of PD. To achieve the goal, a system able to identify the motor status in real time, provide gait guidance system, will be developed. The work presented in this paper will be included as part of it and the data from PD patients that is being gathered [11] will be used to continue the work presented in this paper

The following section will describe the common step length estimators briefly. Adaptation methods for the estimators are also proposed based on the new sensor location in the same section. Section III and IV deal with the experiments and the results with real measured data. Finally, some conclusions are provided.

II. GAIT SPEED AND STRIDE LENGTH ESTIMATION

This section describes six step length and gait speed estimation methods, found in the literature. First 4 methods presented were developed considering the sensor position near L4-L5 position, close to the center of mass (COM) of human body[12]. Method 5 used the sensors of mobile phone and placed it in the pocket of users’ trouser. In method 6, the sensor was placed on the lateral side of waist, same as our proposed position.

A. Method 1:Inverted Pendulum model

The most common approach to measure the average step length is to consider human gait as an inverted pendulum model [12]. Based on this biomechanical model, Zijlstra et al [12] proposed a relationship between the step length and the vertical displacement of the COM of human body which was finally updated by Alvarez et al[13]:

2 √2 (1)

where K is the calibrating constant, l denotes for the leg length of individual from the pelvis (near the COM region) to hill and h denotes the vertical displacement of COM during each step. h can be computed as the range of double integration of vertical accelerations between the instants of two consecutive initial contacts of leg (when the foot strikes the ground) in the direction of forward motion. The calibration constant K is measured for each individual based on the ration of mean real and mean anticipated step length of a course of reference during training phase.

(2)

In the original work, the acceleration, velocity and displacement data were filtered by zero-lag fourth order Butterworth high pass filter at 0.1Hz to avoid drift error

induced by integration. The result following this method is presented as O1 in this paper.

B. Method 2: Double Pendulum Model

Gonzalez et al [14] modified the previously shown inverted pendulum model into a more complex model. They considered the forward displacement of the COM as to be related with the inverted pendulum same as estimator O1 during the swing phase (from left foot-off to left initial contact) and a second different pendulum during the double stance phase (left initial contact to left foot-off). The reason behind considering the swing phase and double stance for left leg was the proposed sensor placement in left lateral side of the waist. The advantage of this method is that is does not need any calibration constant needed to be measured for individual participants.

This method considers the total step length is the summation of the displacement in both stages as follows:

2 √2 (3)

where SLsp and SLdsp are the step length during swing phase and double stance phase, lfoot is the length of individual´s foot and C is the proportional constant.

SLsp is calculated following method 1 and C was determined by considering that the forward displacement during double stance is proportional to the foot length lfoot. C has been fixed as 0.83 by Han et al[15] and 0.67 by Schmid et al. [16] and are presented as estimator O2.1 and O2.2 using these constants.

C. Method 3: Weinberg’s Algorithm

Weinberg [17] considered the stride length as a function of the difference of maximum and minimum vertical acceleration the vertical displacement of the hip during one step. This method again needs the calibration constant K following the same process as method 1 but does not need any integration and, thus, avoids the drift error.

The proposed estimation of the step length is:

SLM3 = (4)

where amax and amin are the maximum and minimum magnitude of vertical acceleration during each step.

The results from this method are mentioned as O3.

D. Method 4: Gait speed from detailed coefficients of wavelet transform

Martin et al [18] proposed a new method to infer the stride speed and from the typical relationship between stride length and speed. They applied the continuous wavelet transform [18] to the acceleration signals from accelerometers and they analyzed the relationship between the kinetic energy of different walking patterns of human and the energies of the wavelet transform detail coefficient to infer the speed.

Steps were recognized by means of two consecutive negative-to-positive transitions of the resultant signals from the wavelet analysis. 4 novel formulations were proposed, which are presented as O4.1, O4.2, O4.3 and O4.4. The benefit of using

this method is that it could be used in different walking speeds and patterns:

∑ (5.1)

∑ (5.2)

∑ (5.3)

∑ (5.4)

where WEdi is the weighted energy of the wavelet transform detailed coefficients at level i from J decomposition levels of wavelet transform with n number coefficient:

√ 1, 2. … 1

√

(6)

(6)

Finally they measured the stride length as:

(7)

E. Method 5:

Bylemans et al[19] developed an empirical solution to estimate the step length using the accelerometers of mobile phone placed in the user´s right trouser pocket . The advantage of this method is that it is orientation free. However, the method is complex. The step length estimation method is as follows:

0.1∆ ∗

. (8)

where aav is the average vertical acceleration between initial contact of ipsilateral foot and contralateral foot and ∆t is the duration of the step in ms and apeak-diff is the range of vertical acceleration during ∆t. K is a calibration constant that is set for individual subject with the following equation

(9)

Following several trials, the authors mentioned that the value of Kdefault was 750 for men and 630 for women. As we have only male volunteers for the experiment process, we considered the Kdefault only for male and presented as O5 in this paper.

F. Method 6: Optimal parameters for Step length

Shin and colleagues [20] proposed a method to estimate the step lengths using optimal parameters based on a linear combination of walking frequency and the variance of the accelerometer signals during one step. The advantage of this method is that it uses a regression model and they could be used in real time.

The method first calculated the magnitude of acceleration by following equation:

(10)

Noise and effect of gravity were reduced from a by implementing sliding window summation technique (SWST)

11

(12)

where w is the window size that need to be less than the duration of each step. In our experiment, w is set to 40 for around 0.2s interval as the operating frequency is 200Hz.

Step was detected from the zero crossing of the acceleration signal. For each detected step the local acceleration variance 2 and walking frequency f were calculated using the following equations

(13)

∑ (14)

where ā is the mean acceleration during one step and n is the number of steps. The step length were then calculated as

∝. . (15)

where α, β are the weights of walking parameters and γ is a constant that are pre-learned. The regression parameter α, β and γ are learned in a training phase similarly to the training of K in method 1. Results of this method are presented as O6 here.

G. Adapted methods to new sensor location

In this study, six step length and gait speed estimators were considered. Among them 4 methods were developed considering the sensor position near COM of human body. These methods were tried to adapt to our system located on left lateral side of waist. But the new position imposes new restrictions that must be taken into account that were not considered in the original methods. Mainly left steps are clearer measured than right steps. To adapt these methods to the new sensor position, we proposed some adaptations.

For method 1 and 3, we adapted the calibration constant for each individual for left step and right step length separately as follows and then estimated the step length which are presented as estimator A1 and A3.

_

_ (16.1)

_

_ (16.2)

For method 2, instead of a fixed value for C, we propose a training phase based on the same original variables used during swing phase. C could be estimated by using following formula and the step length measured with this value are presented as estimator A2.

(17)

For estimator O6, the step length was calculated using the step detection method followed by the authors whereas for

estimator A6 real detected steps from labeled video analysis were used.

III. EXPERIMENTS

A. Data collection

The experiment was conducted using three healthy adult male neither of whom reported difficulties on walking normally. The age of the participants was between 27 and 32 years and height was between 1.65 meters and 1.80 meters.

An inertial system is used to obtain the acceleration measurements. This device was developed by the CETpD laboratory of UPC, Spain and previously used in[2] [11][21]. This single unit portable device consists of a wireless network of inertial sensors with CPU and µSD card and a 1000mAh Li-on battery. All the electronics are encapsulated in a 98mm x 51mm x 17mm white case. The weight of the device including the battery is 125g. The inertial measurement unit of the prototype is composed of a 3-axis accelerometer (LIS3LV02DQ, ±6g range), a 3-axis gyroscope (IDG650+ISZ650, ±20000/s range) and a 3-axis magnetometer (HMC6042+HMC1041Z, ±6G range). A Bluetooth device is also connected with the system for wireless communication. To optimize the energy consumption and control the battery there is also a system included and it interacts with user by three LEDS that indicate the battery level and the current state. The operational frequency of the device is 200Hz. Fig. 1 shows the prototype of the device.

Fig. 1. The inertial system prototype (9x2, Version 6)

In this study, the sensor is placed in a user friendly position [7], the left lateral side of the body near the anterior superior iliac spine (ASIS). The ASIS points are easily found by having the participant trace their fingers along the bony protrusion at either side of their hips. Once they have found the crest (high point) of the ilium bone (the big hip bone), one simply drops below the crest by 2-5 centimeters and aligns the sensor horizontally with that spot. These measures affix the sensors so that they move with the pelvis, rather than with the surrounding soft tissues.

The task that the 3 volunteers performed consisted of a straight walk along a 26m long flat corridor. The participants were asked to stand still for several seconds before starting and then were cued to start. The walk was at the participant’s own “comfortable pace”. When the participant reached the end, he was asked to hold still for a few seconds. The signals during these steady periods (before start and end of the walk) were used later to identify the start and end of walking. The mean acceleration during these periods towards all 3 axis are also used to remove the effect of gravity and to adjust the alignment

of 3 axis of accelerometer by subtracting them from the acceleration signal during walking.

A Casio Exlim high speed video camera is used to record the gait events at 200fps. Following the same methodology as of Sama et al [11], the video recording and the movement signals were synchronized by using a fall of sensor event at the beginning and at the end of each video recording.

B. Data analysis

The recorded video is analyzed by an open-source sport technical analysis software KINOVEA[22] to measure accurate step length, gait speed and travelled distance. Before starting the test and video recording, a sticky paper of 0.16 m length was pasted below the knee on the volunteer’s trouser as a reference length. The video was recorded following the subject from left lateral side parallel to the waist covering the sticky paper and both of the subject’s leg with ground. Using KINOVEA, it is possible to measure the step length by drawing a line between two points where the foot initially touches the ground circled as blue in the fig. 3. But, since video is recorded perpendicular to the forward displacement, measuring length by drawing line between two initial contacts will be biased by step width. In order to reduce the accuracy error, grid lines are used to find the point of two initial contacts in a line. Measurements are obtained through the line between these two points. As the subject and the camera are both moving during the video recording, it is hard to fix the position of the camera with the moving body. Therefore, the measurements are needed be calibrated for every step to get more accurate results. For this reason, a new line is also drawn along the sticky paper in the video to measure the length of it. As the length is already known (0.16 m), it is used to calibrate the measurement of every step length Fig. 2 shows the process to measure one step length with calibration.

Fig. 2: Measuring step length using KINOVEA

An application developed under REMPARK project [11] is used to synchronize the video and the movement signals gathered and to label initial contact and foot-off events on the signals. The program enables the labeling of the gait events manually with the support of video as shown in fig. 3. For every pair of dotted lines, the first one is the initial contact of ipsilateral foot and second one is foot-off of contralateral foot event. Instead of selecting one single point of initial contact and foot-off, periods of start and end time of initial contact and toe-off are labeled and the events are considered in between this periods. During analysis, the first and last 2 steps from

each test are discarded as it is found from the video analysis that the gait is not steady at that time.

Before applying the movement signals, to estimate the step length, they are filtered with zero-lag fourth order low pass Butterworth filter with a cut-off frequency of 30Hz to reduce the noise for all the methods accept method 4.

IV. RESULTS AND DISCUSSION

Table 1 compares the anticipated distance estimated against the real distance during test case, i.e. the second part of the walk, and, additionally, both values for the total walk. Table 2 and 3 show a comparison between real and anticipated average step length and average gait speed, respectively, measured by means of the estimators during the test phase. Standard error of each method were also calculated using Root Mean Square Error (RMSE) and Standard Deviation (SD), which are showed together as RMSE(SD) in the tables. Estimators O1, O2.1, O2.2, O3, O4.1, O4.2, O4.3, O4.4, O5 and O6 follow the original algorithm for the new sensor location. On the other hand, A1, A2,A3, A5

and A6 follow the proposed adaptation methods. As method 4 directly measures the average gait speed and from the speed the average step length, the RMSE and SD could not be calculated for O4.1, O4.2, O4.3 and O4.4.

Results show that all the original methods have a tendency to over-estimate the step length as well as gait speed and distances. O2.1 and O2.2 provides inconsistent results i.e overestimate for one subject and under-estimate for the others. Estimators O4.3, O4.4 and O5, provides the highest error. O1, O3 and O6 provide very good and consistent results i.e low SD and O3 and O6 has the highest performance.

The adapted methods A1, A2, A3, A5 and A6 show significant improvement in their performance through a significant reduction in the RMSE and SD. But still they have the tendency to overestimate. A5 provide the highest error same as the original method O5. With the reduced error and improved result, we found that performance of A6 is the closest to the real value.

TABLE 1: REAL AND ANTICIPATED DISTANCE ESTIMATED BY THE ORIGINAL

(ON) AND THEIR ANTICIPATED (AN) METHODS V1 V2 V3

Total (m)

Test (m)

Total (m)

Test (m)

Total (m)

Test (m)

Real 24.68 12.20 23.42 11.61 23.72 11.76 O1 25.86 13.38 24.50 12.69 24.81 12.84 A1 25.77 13.29 24.47 12.67 24.78 12.81 O2.1 24.12 12.22 31.14 16.04 18.53 9.50 O2.2 22.23 11.28 29.00 14.95 16.71 8.59 A2 25.27 12.79 20.83 10.46 24.41 12.44 O3 25.18 12.70 24.31 12.50 23.91 11.95 A3 25.16 12.68 24.23 12.43 23.88 11.91 O4.1 22.94 11.47 34.38 17.59 23.81 11.90 O4.2 22.94 11.47 34.38 17.59 23.81 11.90 O4.3 82.82 41.41 84.05 43.00 97.59 48.79 O4.4 82.82 41.41 84.05 43.00 97.59 48.79 O5 32.23 16.16 33.62 17.42 34.96 17.45 A5 32.29 16.09 33.61 17.41 34.93 17.42 O6 23.75 11.87 24.25 13.56 23.93 11.97 A6 23.92 11.96 24.22 12.42 23.92 11.96

TABLE 2: REAL AND ANTICIPATED AVERAGE STEP LENGTH DURING TEST

PHASE V1 V2 V3

Test (m)

RMSE(SD) (m)

Test (m)

RMSE(SD) (m)

Test (m) RMSE(SD)

(m) Real 0.64 0.53 0.62

O1 0.70 0.0848

(0.0579) 0.58

0.0712 (0.0515)

0.68 0.0623

(0.0281)

A1 0.70 0.0720

(0.0436) 0.58

0.0641 (0.0426)

0.68 0.0618

(0.0232)

O2.1 0.64 0.0971

(0.1326) 0.73

0.3033 (0.2280)

0.50 0.1403

(0.0773)

O2.2 0.59 0.1076

(0.0973) 0.68

0.2730 (0.2280)

0.45 0.1827

(0.0773)

A2 0.67 0.0960

(0.0973) 0.48

0.2261 (0.2857)

0.65 0.0834

(0.2330)

O3 0.67 0.0491

(0.0412) 0.57

0.0880 (0.0781)

0.63 0.0262

(0.0242)

A3 0.67 0.0465

(0.0389) 0.56

0.0502 (0.0338)

0.63 0.0231

(0.0216) O4.1 0.60 0.80 0.63 O4.2 0.60 0.80 0.63 O4.3 2.18 1.95 2.57 O4.4 2.18 1.95 2.57

O5 0.85 0.2320

(0.1022) 0.79

0.2656 (0.0454)

0.92 0.3137

(0.0836)

A5 0.85 0.2192

(0.0786) 0.79

0.2656 (0.0284)

0.92 0.3072

(0.0668)

O6 0.62 0.0296

(0.0242) 0.57

0.0499 (0.0325)

0.63 0.0232

(0.0212)

A6 0.63 0.0266

(0.0246) 0.53

0.0495 (0.0332)

0.63 0.0232

(0.0211)

Regarding the adaptation methods A1 and A3, the average step length and speed are the same as their original methods O1 and O3. However, considering the variability between left leg and right leg by using a separated calibrating constant for each leg, the errors, both RMSE and SD, are significantly reduced. A2 provides consistent result with better performance than its original methods. The reason is that in the original method, a fixed proportional constant C was considered. In our case, method A2, C is calculated for every subject with the proposed adaptation method and, thus, the performance is improved.

TABLE 3: REAL AND ANTICIPATED AVERAGE STEP LENGTH DURING TEST

PHASE V1 V2 V3 Test RMSE(SD) Test RMSE(SD) Test RMSE(SD)

Fig. 3. Labeled gait event of acceleration signal. LIC and RIC are the left andright initial contact and LFO and RFO are the left and right foot-off event.

(m) (m) (m) (m) (m) (m) Real 1.10 0.87 1.21

O1 1.21 0.1414

(0.0947) 0.95

0.1172 (0.0844)

1.33 0.1225

(0.0550)

A1 1.20 0.1212

(0.0717) 0.95

0.1054 (0.0698)

1.32 0.1220

(0.0467)

O2.1 1.10 0.1685

(0.1707) 1.20

0.5011 (0.3771)

0.98 0.2745

(0.1484)

O2.2 1.02 0.1855

(0.1693) 1.12

0.4507 (0.3764)

0.89 0.3582

(0.1489)

A2 1.16 0.1717

(0.2300) 0.78

0.3695 (0.4643)

1.28 0.1589

(0.4518)

O3 1.15 0.0825

(0.0690) 0.93

0.1443 (0.1277)

1.23 0.0508

(0.0469)

A3 1.14 0.0789

(0.0658) 0.93

0.0815 (0.0546)

1.23 0.0447

(0.0419) O4.1 1.01 1.26 1.21 O4.2 1.01 1.26 1.21 O4.3 3.66 3.09 4.95 O4.4 3.66 3.09 4.95

O5 1.46 0.4064

(0.1860) 1.30

0.4359 (0.0487)

1.80 0.6173

(0.1691)

A5 1.45 0.3833

(0.1458) 1.30

0.4358 (0.0299)

1.80 0.6043

(0.1366)

O6 1.07 0.0507

(0.0414) 0.93

0.0813 (0.0527)

1.23 0.0434

(0.0395)

A6 1.08 0.0455

(0.0419) 0.87

0.0805 (0.0539)

1.23 0.0436

(0.0394)

In method 6, there is significant difference in the performance obtained in the step length estimated using their own step detection algorithm (O6) compared to the estimation obtained by means of the labeled steps. By examining the step detection performed by both methods, it was found that the step detection method accurately detects the left initial contact but over detects right initial contact sometimes during left toe-off as seen in fig. 5. The magenta line in fig. 5 shows magnitude of acceleration where every positive to zero crossings points are detected as steps marked as red circle. The dotted lines are labeled events of ipsilateral and contralateral foot as explained before.

Fig. 4. Step detection using method 6 compared with labeled signal

In method 6, there is significant difference in performance between step length estimated using their own step detection algorithm (O6) and with the labeled steps. We compared the two methods and found that the step detection method accurately detects the left initial contact but detects right initial contact with a delay more precisely during left foot-off.

Though it needs a training phase to calibrate data for every individual, and though the performance is the 2nd best among

the others, we could see that the simplest method is method 3. It also needs less computation compared to the others which make it very suitable to use it in real-time. Method 6 could also be used in real time when the optimal parameters are learned from a training phase.

CONCLUSION

In this study, 6 different methods for step length and gait speed estimation were selected from literature to adapt them to a new sensor location. The original and the adapted methods are compared with real preliminary data. Estimators A2 and A3

show very similar performances and, among them, A3 has the lowest error. Method O6 and A6 provide the closest result to the real data with lowest error. Their low RMS and SD values also show a good significance of their accuracy. Method 6.1 estimates gait speed and step length with the lowest error and shows that an improvement of step detection in method O6 can enhance its performance and can provide the best performance.

This is a preliminary study (with 3 participants). A2, A3, O6 and A6 are those methods that are going to be further explored. The movement signals database that is being gathered from Parkinson Disease patients under REMPARK project [11] will be used to test the enhanced version of the selected methods.

ACKNOWLEDGEMENTS

This work was supported in part by the Erasmus Mundus Joint Doctorate in Interactive and Cognitive Environments, which is funded by the EACEA Agency of the European Commission under EMJD ICE FPA nº 2010-0012.

This work has been performed in the framework of the FP7 project REMPARK ICT- 287677, which is funded by the European Community. The author(s) would like to acknowledge the contributions of their colleagues from REMPARK Consortium (http://www.rempark.eu)

REFERENCES [1] M. M. Hoehn and M. D. Yahr, “Parkinsonism: onset, progression, and

mortality. 1967.,” Neurology, vol. 57, no. 10 Suppl 3, pp. S11–26, Nov. 2001.

[2] A. Samà, C. Angulo, D. Pardo, A. Català, and J. Cabestany, “Analyzing human gait and posture by combining feature selection and kernel methods,” Neurocomputing, vol. 74, no. 16, pp. 2665–2674, Sep. 2011.

[3] A. Sama, C. Perez-Lopez, J. Romagosa, D. Rodriguez-Martin, A. Catala, J. Cabestany, D. A. Perez-Martinez, and A. Rodriguez-Molinero, “Dyskinesia and motor state detection in Parkinson’s Disease patients with a single movement sensor,” in Engineering in Medicine and Biology Society (EMBC), 2012 Annual International Conference of the IEEE, 2012, pp. 1194–1197.

[4] C. G. Goetz, G. T. Stebbins, L. M. Blasucci, and M. S. Grobman, “Efficacy of a patient-training videotape on motor fluctuations for on-off diaries in parkinson’s disease,” Movement disorders, vol. 12, no. 6, pp. 1039–1041, 1997.

[5] J. D. Schaafsma, N. Giladi, Y. Balash, A. L. Bartels, T. Gurevich, and J. M. Hausdorff, “Gait dynamics in Parkinson’s disease: relationship to Parkinsonian features, falls and response to levodopa,” Journal of the neurological sciences, vol. 212, no. 1, pp. 47–53, 2003.

[6] B. Kostek, K. Kaszuba, P. Zwan, P. Robowski, and J. Slawek, “Automatic assessment of the motor state of the Parkinson’s disease patient--a case study,” Diagnostic Pathology, vol. 7, no. 1, p. 18, 2012.

[7] F. Bugané, M. G. Benedetti, G. Casadio, S. Attala, F. Biagi, M. Manca, and A. Leardini, “Estimation of spatial-temporal gait parameters in level walking based on a single accelerometer: Validation on normal subjects

by standard gait analysis,” Computer Methods and Programs in Biomedicine, vol. 108, no. 1, pp. 129–37, Oct. 2012.

[8] E. Martin, V. Shia, and R. Bajcsy, “Determination of a Patient’s Speed and Stride Length Minimizing Hardware Requirements,” in Body Sensor Networks (BSN), 2011 International Conference on, 2011, pp. 144–149.

[9] M. J. Mathie, J. Basilakis, and B. G. Celler, “A system for monitoring posture and physical activity using accelerometers,” Engineering in Medicine and Biology Society, 2001. Proceedings of the 23rd Annual International Conference of the IEEE, vol. 4, pp. 3654–3657, 2001.

[10] REMPARK, “About the project,” 2013. [Online]. Available: http://www.rempark.eu/. [Accessed: 12-Apr-2013].

[11] A. Samà, C. Perez, D. Rodriguez-Martin, J. Cabestany, J. M. Moreno Aróstegui, and A. Rodríguez-Molinero, “A Heterogeneous Database for Movement Knowledge Extraction in Parkinson’s Disease,” in European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, 2013.

[12] W. Zijlstra and A. L. Hof, “Displacement of the pelvis during human walking: experimental data and model predictions,” Gait & posture, vol. 6, no. 3, pp. 249–262, 1997.

[13] D. Alvarez, R. C. González, A. López, and J. C. Alvarez, “Comparison of step length estimators from weareable accelerometer devices,” in Engineering in Medicine and Biology Society, 2006. EMBS’06. 28th Annual International Conference of the IEEE, 2006, pp. 5964–5967.

[14] R. C. González, D. Alvarez, A. M. López, and J. C. Alvarez., “Modified pendulum model for mean step length estimation,” in Engineering in Medicine and Biology Society, 2007. EMBS 2007. 29th Annual International Conference of the IEEE., 2007, vol. 2007, pp. 1371–4.

[15] T. R. Han, N. J. Paik, and M. S. Im, “Quantification of the path of center of pressure (COP) using an F-scan in-shoe transducer.,” Gait & posture, vol. 10, no. 3, pp. 248–54, Dec. 1999.

[16] M. Schmid, G. Beltrami, D. Zambarbieri, and G. Verni, “Centre of pressure displacements in trans-femoral amputees during gait.,” Gait & posture, vol. 21, no. 3, pp. 255–62, Apr. 2005.

[17] H. Weinberg, “Using the ADXL202 in pedometer and personal navigation applications,” Analog Devices AN-602 application note, 2002.

[18] E. Martin, “Novel method for stride length estimation with body area network accelerometers,” 2011 IEEE Topical Conference on Biomedical Wireless Technologies, Networks, and Sensing Systems, pp. 79–82, Jan. 2011.

[19] I. Bylemans, M. Weyn, and M. Klepal, “Mobile phone-based displacement estimation for opportunistic localisation systems,” in Mobile Ubiquitous Computing, Systems, Services and Technologies, 2009. UBICOMM’09. Third International Conference on, 2009, pp. 113–118.

[20] S. H. Shin and C. G. Park, “Adaptive step length estimation algorithm using optimal parameters and movement status awareness,” Medical engineering & physics, vol. 33, no. 9, pp. 1064–1071, 2011.

[21] A. Sama, A. Catala, A. Rodr\iguez-Molinero, and C. Angulo, “Extracting Gait Spatiotemporal Properties from Parkinson`s Disease Patients,” in Neural Information Processing Systems, 2009.

[22] “Kinovea, A microscope for your videos.” [Online]. Available: http://www.kinovea.org/. [Accessed: 12-Apr-2013].