Over the years, several UWB localization systems have already been proposed for accurate position estimation of pedestrians. Most of these systems already proposed for pedestrian localization have only been individually evaluated for a particular wearable sensor position. In this paper, we compare the positioning performance of seven body wearable sensor positions i.e., chest, arm, ankle, wrist, thigh, fore-head, and hand using an Extended Kalman Filter (EKF) based localization algorithm in an indoor environment. The conclusion drawn is that the fore-head is the best, and the chest is the worst body sensor location for tracking a pedestrian. While the fore-head position is able to set an error lower than 0.35 m (90th percentile), the chest is able to set 4 m. The fore-head's performance is followed by the hand, ankle, wrist, thigh, arm, and chest in that order.
Wearable connected devices hold a great deal of promise, and therefore their market is expected
to reach $ 70 billion in 2025 [
Indoor localization of pedestrians is still an open problem though many different approaches
have been proposed using diverse technologies to obtain a similar performance to the Global
Navigation Satelite System (GNSS) outdoors. For instance, Wi-Fi, Ultrawideband (UWB), or
Ultrasound, and advanced processing techniques, such as Kalman and Particle Filters, have been
proposed to cope with the deterioration in performance due to combined effects of pathloss
and multipath fading. Among the most accurate localization solutions are those that rely on
ultrasound or UWB radio signals but both their performances deteriorate in non-line-of sight
(NLOS). UWB signals can achieve decimeter-level errors in indoor positioning and has gained
significant interest in research as a promising candidate for industry solutions. In recent years,
there has been a great deal of interest in its use for positioning [
However, in tracking context of pedestrians using UWB technology, an important factor which has been often overlooked but has significant effect on the positioning error is the influence caused by the human body itself. The effects of human body shadowing are additional propagation losses or biases in time of flight (TOF) measurements when the body blocks the line-of-sight (LOS) between a wearable sensor and an anchor. These effects, which generally depend on: i) body wearable sensor position, and ii) relative heading angle are currently, not adequately accounted for and will decrease the accuracy of localization systems.
In recent years, a few works have focused on the ranging and positioning error introduced
by the human body due to the relative heading angle, leaving aside the impact of body
wearable sensor positions [
The rest of this paper is organized as follows: experimental setup and the method adopted to collect the experimental data is described in Section II. The UWB positioning performance is analyzed in sections III for different body wearable sensor positions. In section IV, the presentation and discussion of results is made. Finally, in the last section, we give some conclusions and future work. 2
Throughout this paper, several TREK1000 development kits manufactured by Decawave were
used. According to [
The experiments were carried out inside the Luis Mercader Lab at the department of Electric, Electronic and Communication engineering at the Public University of Navarra in Spain. The Lab had the following dimensions: 6 m wide, 13 m long, and 4 m high, and contained a number of computers, monitors, chairs, desks, closets and working people. The floor and ceiling were made of concrete. The floor plan of lab environment where the tests were performed is showed in Fig. 1.
The floor plan shows detailed anchor positions and a path with 26 ground-truth points with each point approximately 1 m from the other. There are several interfering objects such as pieces of furniture, metallic cabinets and desktop computers. The origin of the reference system is defined at the top left corner. The UWB anchors were mounted on tripods in the positions indicated in Table 1. 1240 1240 548 Anchor 2 6 7 8
A male subject, 1.80 m height and 77 kg mass was considered for the measurements. The wearable sensors were mounted on the subject with the help of Velcro straps at the right-ankle, right-thigh, fore-head, right-hand, right-arm, chest, and right- wrist as seen in Fig. 2.The heights at which the wearable sensors were mounted are showed in Table 2.
Using the installed ground marks, firstly, the wearable sensor was mounted on a tripod and moved along the path starting from test point 1 and ending at test point 26 [see – Fig. 1] as a reference for further comparisons.
Similar to the work in [
Because the main application of ranging is positioning, in this section we compare the positioning performance. As explained in the previous section, the infrastructure of the experiments consists of one UWB node mounted on seven body locations and four UWB nodes in fixed and known positions acting as anchors. In order to achieve the positioning performance, a localization algorithm based on the EKF is implemented. We use range measurements from the continuous scenario. To minimise the effects of NLOS, the measured ranges with a very large error were assumed to be under NLOS conditions. And so, they were rejected in the update of the state vector estimation. (a) Chest (b) Arm (c) Ankle (d) Wrist (e) Thigh (f) Fore-head (g) Hand
Similar to [
The state vector xk is defined by the 3-D position (p) and velocity (v) estimates as indicated in (1) as :
Therefore, xk is defined as : where F k is the state transition matrix given as :
xkjk = xk = (pk vk )T xkjk 1 = F kxk 1jk 1 + wk
01 0 01 F k = @0 1 0A
0 0 1 0 2 ax
T 2=2 where T is equal to the time difference between timestamps k and k 1 and ax = 100 cm, ay = 100 cm , and az = 10 cm are the uncertainty that model the acceleration driving noise of the dynamic model in the x, y, and z directions. The state vector xk is initialized with the coordinates of the first ground marks. Their initial standard deviations were set to 20 cm for the x and y coordinates, and 0.5 cm for the z coordinate. These parameters were set from preliminary measurements to characterize the UWB nodes and the motion of the subject.
The measurement model has the form:
zk = h(xkjk) + nk where zk is the measurements vector, h is the measurement non-linear function, and nk is the measurement noise with covariance matrix Rk.
The standard deviation (SD) of the measurement model was set according to Table 3. The values were obtained from a set of measurements to characterize the UWB nodes for each wearable sensor position.
Using UWB ranges from the 4 anchors to update the estimate of the state xkjk, the measurements take on the following form:
zn;k = hn(xkjk) = = q (px ax;n)2 + (py ay;n)2 + (pz az;n)2 where zn;k is the measured range between the nth anchor at the position ax;n; ay;n, az;n, and wearable sensor with current position estimates at px; py and pz. (6)
Because the output of the EKF filter provides a continuous estimation of the position, it is necessary to detect when the subject feet are still so that the ground truth positions are estimated. Therefore, we performed the following tasks for each of the body wearable positions. 1. Smoothing: where a moving average filter is applied to each position component. The length of the window was empirically set to 4 s. 2. Filtering: in order to reduce the number of outliers in the position estimation, a moving variance with window length of 4 s was also set. 3. To determine the ground truth position, the k-means clustering is applied to the filtered estimates. The ground truth positions are extorted from the obtained centroids. 4
To compare the performance of the different wearable sensor positions, the absolute value of the difference between the real ground truth positions and the values estimated by the EKF are used as the error metric. To better see the performance results, the main statistics are presented in the Table 4. In Table 4, we differentiate the performance results according to the sensors that worn on the the frontal plane (fore-head, chest and hand) and the side plane (ankle, wrist, thigh, and arm) of the user.
Looking at Table 4, the first observation is the existence of a clear relationship between the localization performance and the body wearable sensor position. Overall, it can be observed that the fore-head and the chest position give the best and worst possible localization performance, respectively. In absolute terms, 90 % of the estimates were below the error of 0.35 m and 4.04 m for the fore-head and chest, respectively.
The performance of the chest is heavily influenced by NLOS conditions due to human body shadowing. Though all the wearable positions are influenced by human body shadowing effect, this effect is more pronounced at the chest because the large size of the chest allows a lot more power to get absorbed by the body, which explicitly translates in to extremely large errors. Therefore, the chest position can be used if it is possible to install enough anchors to minimize the risk of NLOS.
Arm
After the fore-head, the hand position obtains the second lowest errors as 90 % of the estimates were below the error of 0.62 m. The hand performs better than the chest (though in both cases the wearable is at the center of the chest) because the space of 20 cm [see – Fig. 2g] allows for creeping wave propagation during NLOS. Among the wearable positions where the sensor is placed on side plane of the user, the ankle gives the best performance as 90 % of the estimates were below the error of 0.97 m. Following the performance of the ankle position is the wrist, thigh, and arm locations with 90th percentile of 1.14 m, 1.46 m, 2.47 m, respectively. Similar to the ranging performance, the differences in the position errors among these locations can be attributed to the differences in the height and size of the limb on which the sensor is attached. 5
We have presented an experimental comparison of seven body wearable sensor positions available for tracking the position of a pedestrian. The comparison has been done in an indoor environment with LOS and NLOS conditions. Several NLOS have been created by the body of the user which acts as an obstacle of the direct signal path between the wearable sensor and the access point. While the performance of forehead has been superior, the performance results for the chest position has been observed to be lower than the other wearable sensor positions.