Localization of Moving Objects Based on RFID Tag Array and Laser Ranging Information

Abstract

RFID (radio-frequency identification) technology is rapidly emerging for the localization of moving objects and humans. Due to the blockage of radio signals by the human body, the localization accuracy achieved with a single tag is not satisfactory. This paper proposes a method based on an RFID tag array and laser ranging information to address the localization of live moving objects such as humans or animals. We equipped a human with a tag array and calculated the phase-based radial velocity of every tag. The laser information was, first, clustered through the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm and then laser-based radial velocity was calculated. This velocity was matched with phase-based radial velocity to get best matching clusters. A particle filter was used to localize the moving human by fusing the matching results of both velocities. Experiments were conducted by using a SCITOS G5 service robot. The results verified the feasibility of our approach and proved that our approach significantly increases localization accuracy by up to 25% compared to a single tag approach.

1. Introduction

In the last few years, there has been significant research in the area of tracking, navigation, and localization of objects in different environments. Outdoor localization is easy to manage with the help of GPS, while, indoor localization has a lot of challenges like environmental factors (humidity, thickness of walls, and temperature) that can greatly affect sensor feedback, which ultimately decreases the localization accuracy. A typical indoor localization system may use different technologies for localization, for example UWB (ultra-wideband) [1], Wifi [2], and RFID (radio frequency identification) [3]. Every technology has its limitations and advantages; for instance, printed bar codes are typically read by an optical scanner that requires a direct line-of-sight to detect and extract information [4]. On the other hand, RFID technology is easy to use due to its unique qualities, such as a wide range of tag types including passive and active tags, identification through an unique ID [5], contactless feature [6], small size [7], and cost effectiveness [8]. Hence, RFID is not only a preferred technology for indoor localization but also the most convenient one. RFID technology can facilitate indoor localization in different complex environments, such as in the areas of supply chain management, inventory control, and human or robot localization, in a cost effective way [9].

Researchers have developed different localization methods to estimate the distance of the objects by using RSS and phase information. Ma et al. [10] proposed a method to estimate the position of a mobile object using RSS and phase information. Liang et al. introduced a device-free indoor localization system based on particle swarm optimization which uses RSS and phase information measured by RFID readers to localize the target [11]. Ruan et al. proposed a data driven approach for human localization and tracking [12] and previously, they also used RSS in another research [13] and achieved a reasonable level of accuracy in both approaches; however, RSS is easily affected by environmental noise and other interferences. RFID phase readings are also used for localization but these methods provide less accuracy due to phase ambiguity. Some researchers have tried to reduce phase ambiguity, e.g., Liu et al. has tried to reduce ambiguity of phase readings with comparatively less computational requirements [14]. It is known that radio signals are always vulnerable to environmental factors like object type, surface patterns, temperature, and humidity. Han et al. tried to increase the reading range in the specific environment of high humidity by attaching a 3D printed grapheme antenna to RFID tags [15]. In some cases, it may also be required to localize objects at a large distance which may be outside the reading range of RFID readers. Some researchers have also improved localization accuracy at blind locations which can occur in when the location is outside the range of the RFID reader’s antenna beam [16]. Accuracy is always limited at blind locations as the methods try to predict the location of the target from its previous location inside of the antenna range.

In this paper, we localized a moving object encircled with an RFID tag array. As the size, shape and types of moving objects can be different; it is proposed that a tag array be used instead of a single tag for the localization of moving objects such as humans. A live object, such as a human, may change its location frequently in a short time so encircling the human body with an RFID tag array is helpful to localizing a human body more accurately. RFID tags are cheap in cost so it is convenient to use an RFID tag array instead of a single tag. RFID tags are small in size and weigh very little [17] so it is easy to attach an RFID tag array on clothes or belts. This approach has an ease of use as the implementation of this approach does not need any training or reference tags and it also does not require any complex setup of antennas. This approach significantly increases the performance of the indoor localization of moving objects and also allows the objects to move freely in the environment.

Subsequent sections of the paper are organized as follows. Related work is given in Section 2. System details are described in Section 3 including tag array approach, particle filter and laser clustering techniques. In Section 4, experiment details and results are discussed. Conclusions and possible future extensions are described in Section 5.

2. Related work

This section gives a detailed view of literature about different technologies and methods proposed and implemented for indoor localization. The related work is divided into the following three categories:

• RFID-based Systems: In RFID-based systems, time of arrival (TOA), time difference of arrival (TDOA), RSS, and angle of arrival (AOA) are the most commonly used models. Shen et al. introduced “ANTspin” which uses the AOA technique for the localization of a static tagged object by spinning antennas [18] while Abkari et al. combined TOA and AOA techniques for localization in a hospital environment [19]. Ma et al. proposed a tag-to-tag communication system to achieve multi-tag cooperative localization using the TDOA method [20]. Duan et al. used a SAR (synthetic aperture radar)-based model to estimate the spinning angle of RFID tags and achieved high accuracy [21] but spinning the tags is not possible in every scenario. Some researchers used RSS-based algorithms for localization, e.g. Kaltiokallio et al. used RSS in a recursive particle filter to achieve a high localization accuracy, but it also required a high number of reference tags [22]. Hsiao et al. used the “Real Time Location System” with multiple reference tags to enhance the localization accuracy of a person equipped with RFID tags [23]. Zhang et al. [24] used RSS in k-nearest neighbor (k-NN) algorithm and tried to reduce localization error. Zhao et al. proposed a range-based method using RSS and analyzed the similarity of backscattering; this method needs reference tags in the environment to get better localization results [25]. Wu et al. proposed an unwrapped phase position model to localize a static RFID tag [26]. Xiao et al. attached two tags to the same item and exploited the RFID phase for the localization of relatively small items in a controlled environment [27]. Buffi et al. tested and validated a localization method for tagged objects moving along a conveyor belt, where, in their experiments, the speed and path of the objects was already known [28]; however, tracking a human being is more complex as the movement path and speed of humans and animals are random. In another method [29], Buffi et al. achieved a high localization accuracy for static tags by using the SAR (synthetic aperture radar) approach in the context of the RFID phase-based technique.

• Sensor Fusion based Systems: Researchers fused information from different sensors (e.g., laser, vision sensors, sound etc.) for the localization of the objects. Liu et al. combined RFID and a 2d laser range finder to follow a dynamic RFID tag with obstacle avoidance capabilities using a mobile robot [30]. Wu et al. used a fuzzy reasoning algorithm, a fused laser and RFID to localize a robot using a single RFID tag [31]. Duan et al. proposed to fuse vision and RFID sensors to get a high accuracy [32] but because of the limitations of vision sensors, it could only work in specific environments. Llorca et al. implemented a people localization system for outdoor environments by fusing RFID and BLE (Bluetooth Low Energy) with a vision system and was able to recognize individuals in a group [33]. Alfian et al. proposed an RFID-based tracking system for food items by utilizing data mining techniques to predict missing sensor data [34]. Yang et al. improved particle filters by using some characteristics of moving direction in a densely tagged environment, but this method needs a much longer time to build the system model [35]. Wu et al. proposed another method to localize by the fusion of RFID and inertial measurement units (IMU) [36]. They created a virtual map with the help of four persons carrying RFID tags and compared the map results with RSS to increase localization accuracy. Proximity-based approaches were also tested by different researchers for localization in RFID systems. For example, Sequeira et al. constructed a model to localize moving humans wearing RFID tags and tried to build proximity between the robot and human [37]; however, proximity-based approaches are not cost-effective as an array of antenna is required for implementation.

• Hybrid Systems:RFID technology is also combined with other positioning systems such as GPS, Wifi, ZigBee and Bluetooth for localization. Suparyanto et al. combined inertial sensors and iGPS with RFID and developed a position monitoring system to localize container trucks [38]. Digiampaolo et al. presented a global localization system for an indoor autonomous vehicle equipped with odometer and RFID reader [39]. Zheng et al. also increased localization accuracy by combining GPS and RFID technologies [40]. Some researchers merged different technologies to improve the accuracy of indoor localization; for example, Wu et al. combined MM (magnetic matching), PDR (Pedestrian Dead Reckoning) and RFID in their research [41]. Recently, Luo et al. used Wifi and a RFID tag array to localize an educational robot [42], but arranging Wifi devices in a specific layout is needed in this hybrid system.

The different approaches are compared in Table 1.

Existing approaches require reference tags, antenna arrays, or are developed for specific environments/objects, while our approach localizes a human in a cost-effective way with less computation and relatively high accuracy. Details of all the symbols used in this paper are defined in Table 2.

3. System Details

In this paper, we propose a tag array approach to deal with the localization and tracking of live moving objects (human). Previously, Fu et al. had developed an approach to track a moving object by using a single tag carried by a human [43] but we have used a tag array from a different perspective and used DBSCAN for better clustering results of laser data. As the system is supposed to localize the tags at specific timestamps, so the tag array approach will always give better localization results than a single tag approach. We have conducted different experiments to verify the localization accuracy of our approach. The results of our experiments strengthened and proved this argument. Our approach allows the human to move freely in the environment so this approach can satisfy many industrial needs such as tracking patients in hospital environment, monitoring pets, and the localization of a human in a building etc. Secondly, RFID tags are not expensive so it is also a cost effective approach to use tag array instead of only one tag.

As shown in Figure 1, the entire localization system can be divided into three parts. The first part uses an RFID reader and 2d laser range finder to receive the measurements from the environment. The second part processes the measurements collected from the RFID and the laser range finder. The third part uses a particle filter to achieve the tracking and localization of a dynamic object.

Our proposed approach fuses laser and RFID phase information in a particle filter. Laser information is fragmented into different clusters. The laser cluster velocity and phase-based velocity of each tag is calculated to match both velocities to find the position of the tag. Laser raw data are processed with the help of the DBSCAN algorithm to find laser clusters and then calculate laser velocity. At the same time, RFID information which consists of antenna ID, tag ID, phase, and timestamp is collected. Due to multiple tag readings at the same time stamp, we pulled together all tag information to calculate the phase-based RFID velocity of each tag at specific timestamps. We compared laser velocity and phase-based velocity and extracted the similarities between them. Among similarity results, we choose the best laser clusters for the data input of particle filtering. This simple approach highly increases the localization accuracy.

3.1. Tag Array Importance in Our Approach

Some researchers have tried to localize daily used objects (static) which are equipped with a tag array in a limited and controlled environment [44]. Tracking an animal or human in a cost-effective manner in a complex environment still needs a lot of effort. Wang et al. tried to improve the localization accuracy of a moving finger to improve gesture learning through an RFID tag array [45]; however, it cannot be used for relatively large objects like humans. Nakamori et al. read body posture and the probability among deducted objects to be a human body by using laser range finders [46] but only for a specific game-playing environment.

Another group of researchers used a tag array approach in the traditional way of putting the tags side by side and at very short distances [47]. This kind of arrangement may be fine for objects with orthogonal surfaces but for objects like humans or animals, it is better to encircle the body with a tag array. The shape and movements of a human body, like other live objects, is complex due to the rapid and unexpected location change of body parts. If a human is equipped with one RFID tag then it is difficult to calculate the position of the human, while, on the other hand, if a human is equipped with a tag array then rapid changes of human movement can be recorded and localized easily. We have placed the tags on front, back, and at both hands of a human, and used this tag array to represent the object to be tracked, as illustrated in Figure 2.

3.2. Particle Filtering

Particle filter implements a Bayesian framework. The Bayesian inference estimates the probability density function over the state by the sensor input of previous time and the object’s motion information. Particle filter is famous for the solution of nonlinear problems and can approximate posterior probability over a state by using a number of particles. The motion information of the object is utilized to predict the target position at a given time so RFID and laser ranging measurement are used in our experiments to estimate the state and position of an object (i.e., posterior probability):

(1)$\mathrm{p}\left({\mathrm{X}}_{\mathrm{t}}|{\mathrm{g}}_{1:\mathrm{t}},{\mathrm{r}}_{1:\mathrm{t}},{\mathrm{u}}_{1:\mathrm{t}}\right)={\mathsf{\eta }}_{\mathrm{t}}.\mathrm{p}\left({\mathrm{X}}_{\mathrm{t}}|{\mathrm{X}}_{\mathrm{t}-1},{\mathrm{u}}_{\mathrm{t}}\right).{{\displaystyle \sum }}_{\mathrm{tag}=1}^{\mathrm{m}}\mathrm{p}\left({\mathrm{g}}_{\mathrm{t}}|{\mathrm{X}}_{\mathrm{t}},{\mathrm{r}}_{\mathrm{t}}^{\mathrm{tag}}\right).\mathrm{p}\left({\mathrm{X}}_{\mathrm{t}-1}|{\mathrm{g}}_{1:\mathrm{t}-1},{\mathrm{r}}_{1:\mathrm{t}-1},{\mathrm{u}}_{1:\mathrm{t}-1}\right)$

In Equation 1, ${\mathrm{X}}_{\mathrm{t}}$ is the position of the object at time t, ${\mathrm{g}}_{\mathrm{t}}$ is the measurement of the laser range finder at time t, ${\mathrm{r}}_{\mathrm{t}}^{\mathrm{tag}}$ is the measurement of RFID at time t, ${\mathrm{u}}_{\mathrm{t}}$ is the motion information of the object at time t, and ${\mathsf{\eta }}_{\mathrm{t}}$ is a normalizing factor. While $\mathrm{p}\left({\mathrm{X}}_{\mathrm{t}}|{\mathrm{X}}_{\mathrm{t}-1},{\mathrm{u}}_{\mathrm{t}}\right)$ is the motion model, which is used to predict the object position at time t given the previous position ${\mathrm{X}}_{\mathrm{t}}$ and motion information ${\mathrm{u}}_{\mathrm{t}}$. Observation model $p\left({\mathrm{g}}_{\mathrm{t}}|{\mathrm{X}}_{\mathrm{t}},{\mathrm{r}}_{\mathrm{t}}^{\mathrm{tag}}\right)$ describes the likelihood of receiving a measurement ${\mathrm{g}}_{\mathrm{t}}$ (i.e., laser-based clusters) given the RFID measurement ${\mathrm{r}}_{\mathrm{t}}^{\mathrm{tag}}$ and the current state ${\mathrm{X}}_{\mathrm{t}}$. Further, $\mathrm{p}\left({\mathrm{X}}_{\mathrm{t}-1}|{\mathrm{g}}_{1:\mathrm{t}-1},{\mathrm{r}}_{1:\mathrm{t}-1},{\mathrm{u}}_{1:\mathrm{t}-1}\right)$ is the state at time t − 1. Location of a particle ${\mathrm{X}}_{\mathrm{t}}$ is denoted as

(2)${\mathrm{X}}_{\mathrm{t}}={\left\{{\mathrm{X}}_{\mathrm{t}}^{\left[\mathrm{n}\right]},{\mathsf{\omega }}_{\mathrm{t}}^{\left[\mathrm{n}\right]}\right\}}_{\mathrm{n}=1}^{\mathrm{N}}$

where ${\omega }_t^{\left[n\right]}$ is particle weight [2] and ${\mathrm{X}}_{\mathrm{t}}^{\left[\mathrm{n}\right]}$ is represented as

(3) ${\mathrm{X}}_{\mathrm{t}}^{\left[\mathrm{n}\right]}=\left\{x_{\mathrm{t}}^{\left[\mathrm{n}\right]},y_{\mathrm{t}}^{\left[\mathrm{n}\right]}\right\}$

Equation 3 denotes the 2D location of the particle. The particle filter performs three steps: Prediction, update, and resampling. At the prediction stage, we predict the state of a particle $x_t^{\left[n\right]}$. Prediction is computed by using the information of previous state $x_{t-1}^{\left[n\right]}$ and the object’s motion $u_t$. As we do not know the moving direction of the particle so it can be described as

(4)$\{\begin{array}{c}{x}_{\mathrm{t}}^{\left[\mathrm{n}\right]}= x_{\mathrm{t}-1}^{\left[\mathrm{n}\right]}+ N\left(\mathrm{O},{\mathsf{\sigma }}_{\mathrm{t}}\right)\\ y_{\mathrm{t}}^{\left[\mathrm{n}\right]}= y_{\mathrm{t}-1}^{\left[\mathrm{n}\right]}+ N\left(\mathrm{O},{\mathsf{\sigma }}_{\mathrm{t}}\right)\end{array}$

where ${\sigma }_t$ represents the Gaussian noise. We, first, find the nearest cluster $\tilde{l}$ of particle n, and use $\Delta t.v_t^{\tilde{l}}$ as ${\sigma }_t$, where $\Delta t$ is the time difference between two timestamps and $v_t^{\tilde{l}}$ is the laser-based radial velocity of nearest cluster $\tilde{l}$ at time t. At the update stage of the particle filter, we change the previous prediction by using current information and update the weights of the particles. We represent the object location by a number of particles with different weights. Weight ${\mathsf{\omega }}_{\mathrm{t}}^{\left[\mathrm{n}\right]}$ of the particle ${\mathrm{X}}_{\mathrm{t}}^{\left[\mathrm{n}\right]}$ is computed in Equation 5 based on the observation model $\mathrm{p}\left({\mathrm{g}}_{\mathrm{t}}|{\mathrm{X}}_{\mathrm{t}},{\mathrm{r}}_{\mathrm{t}}^{\mathrm{tag}}\right)$.

(5) ${\mathsf{\omega }}_{\mathrm{t}}^{\left[\mathrm{n}\right]}={\mathsf{\eta }}_{\mathrm{t}}.{\mathsf{\omega }}_{\mathrm{t}-1}^{\left[\mathrm{n}\right]}.\mathrm{p}\left({\mathrm{g}}_{\mathrm{t}}|{\mathrm{X}}_{\mathrm{t}}^{\left[\mathrm{n}\right]},{\mathrm{r}}_{\mathrm{t}}^{\mathrm{tag}}\right)$

While combined with the weighted K most similar clusters, the observation model $\mathrm{p}\left({\mathrm{g}}_{\mathrm{t}}|{\mathrm{X}}_{\mathrm{t}},{\mathrm{r}}_{\mathrm{t}}^{\mathrm{tag}}\right)$ is approximated as

(6) $\mathrm{p}\left({\mathrm{g}}_{\mathrm{t}}|{\mathrm{X}}_{\mathrm{t}},{\mathrm{r}}_{\mathrm{t}}^{\mathrm{tag}}\right)={\displaystyle \sum }_{\mathrm{i}=1}^{\mathrm{k}}\mathrm{sim}\left({\mathrm{v}}_{\mathrm{t}}^{\mathrm{i}},{\mathrm{v}}_{\mathrm{t}}^{\mathrm{r}}\right).\exp \left(-\frac{{\mathrm{d}}^2\left({\mathrm{X}}_{\mathrm{t}}^{\left[\mathrm{n}\right]},{\mathrm{C}}_{\mathrm{t}}^{\mathrm{i}}\right)}{2}\right)$

(7) ${\mathrm{d}}^2\left({\mathrm{X}}_{\mathrm{t}}^{\left[\mathrm{n}\right]},{\mathrm{C}}_{\mathrm{t}}^{\mathrm{i}}\right)=\frac{{\left({\mathrm{x}}_{\mathrm{t}}^{\left[\mathrm{n}\right]}-{\overline{\mathrm{x}}}_{\mathrm{t}}^{\mathrm{i}}\right)}^2}{{\mathsf{\sigma }}_{\mathrm{d}}}+\frac{{\left({\mathrm{y}}_{\mathrm{t}}^{\left[\mathrm{n}\right]}-{\overline{\mathrm{x}}}_{\mathrm{t}}^{\mathrm{i}}\right)}^2}{{\mathsf{\sigma }}_{\mathrm{d}}}$

In Equation 7, ${\mathsf{\sigma }}_{\mathrm{d}}$ is translation coefficient to the distance function. An important stage of the particle filter is to resample the particles as per the weight of the particle. Resampling is done when weight of particles is updated. After a reasonable number of iterations, less weighted particles are replaced with a high weighted set of particles.

3.3. Laser Clustering

We use a 2D laser range finder to scan the environment. Laser information is used to get the distance of a moving object. Laser data are divided into different clusters. A clustering algorithm DBSCAN is used for the clustering of laser data. DBSCAN groups together those points which are close to each other based on a distance measurement (usually Euclidean distance). It also marks the points that are in low-density regions as outliers. It uses two basic parameters, Epsilon ε and MinPoints ζ. Different details of these parameters are given as below.

• Epsilon ε is the radius of neighborhood around a point. It means that if the distance between two points is lower or equal to ε, these points are considered as neighbors. If the ε value chosen is too small, a large part of the data will not be clustered. Those values will be considered as outliers because they do not satisfy the number of points to create a dense region. On the other hand, if the value is chosen to be too high, clusters will be merged and the majority of objects will be in the same cluster. The ε should be chosen based on the distance of the dataset but, in general, small ε values are preferable.

• MinPoints ζ is the minimum number of points to form a dense region. For example, if we set the ζ as 5, then we need at least 5 points to form a dense region. As a general rule, a minimum ζ can be derived from a number of dimensions (D) in the data set, as ζ ≥ D + 1. Larger values are usually better for data sets with noise and will form more significant clusters. The minimum value for the ζ must be 2, but for a larger data set, a larger ζ value should be chosen.

Figure 3; Figure 4 represent raw laser data and clustering results using DBSCAN parameters, epsilon ε, and MinPoints ζ.

3.4. Comparison of Phase-based Radial Velocity and Cluster-based Radial Velocity

First, we have to find distance ${\mathrm{d}}_{\mathrm{t}}^{\mathrm{i}}$ between two clusters to find laser-based radial velocity $v_t^i$ and for every cluster $C_t^i$ at time t, we find the nearest cluster $C_{t-1}^{\widehat{j}}$ at the previous time t–1. Therefore, both clusters $\left(C_t^i, C_{t-1}^{\widehat{j}}\right)$ could be considered as the same object at two sequent timestamps. Laser-based radial velocity ${\mathrm{v}}_{\mathrm{t}}^{\mathrm{i}}$ is calculated in Equation 8:

(8)${\mathrm{v}}_{\mathrm{t}}^{\mathrm{i}}=\frac{\Delta {\mathrm{d}}_{\mathrm{t}}^{\mathrm{i}}}{\Delta \mathrm{t}}=\frac{\sqrt{{\left({\overline{\mathrm{x}}}_{\mathrm{t}}^{\mathrm{i}}\right)}^2+{\left({\overline{\mathrm{y}}}_{\mathrm{t}}^{\mathrm{i}}\right)}^2}-\sqrt{{\left({\overline{\mathrm{x}}}_{\mathrm{t}-1}^{\hat{\mathrm{j}}}\right)}^2+{\left({\overline{\mathrm{y}}}_{\mathrm{t}-1}^{\hat{\mathrm{j}}}\right)}^2}}{\Delta \mathrm{t}}$

where (${\overline{\mathrm{x}}}_{\mathrm{t}}^{\mathrm{i}}$, ${\overline{\mathrm{y}}}_{\mathrm{t}}^{\mathrm{i}}$ ) is center of cluster $C_t^i$ and (${\overline{\mathrm{x}}}_{\mathrm{t}-1}^{\hat{\mathrm{j}}}$, ${\overline{\mathrm{y}}}_{\mathrm{t}-1}^{\hat{\mathrm{j}}}$) is center of cluster $C_{t-1}^{\widehat{j}}$. While $\widehat{j}$ is calculated as

(9)$\widehat{j}= \underset{j}{\mathrm{argmin}}\sqrt{{\left({\overline{x}}_t^i- {\overline{x}}_{t-1}^{\hat{\mathrm{j}}}\right)}^2+{\left({\overline{y}}_t^i- {\overline{y}}_{t-1}^{\hat{\mathrm{j}}}\right)}^2}$

where 1 ≤ j ≤ Nt−1 and radial velocity of the tag based on phase difference ${\mathrm{v}}_{\mathrm{t}}^{\mathrm{r}}$ is calculated as

(10) ${\mathrm{v}}_{\mathrm{t}}^{\mathrm{r}}=\frac{{\mathsf{\Delta }\mathsf{\phi }}_{\mathrm{t}}}{\mathsf{\Delta }\mathrm{t}}=\frac{\mathrm{c}}{4\mathsf{\pi }\mathrm{f}\mathsf{\Delta }\mathrm{t}}.\left({\mathsf{\phi }}_{\mathrm{t}}-{\mathsf{\phi }}_{\mathrm{t}-1}\right)$

In Equation 10, $\mathrm{c}$ is the velocity of light,$\mathrm{f}$ is signal frequency, and ${\mathsf{\Delta }\mathsf{\phi }}_{\mathrm{t}}$ is the phase difference of the tag at two successive times. Comparison results derived from Equation 11 can identify whether both of them are from the same object or not.

(11) $\mathrm{sim}\left({\mathrm{v}}_{\mathrm{t}}^{\mathrm{i}},{\mathrm{v}}_{\mathrm{t}}^{\mathrm{r}}\right)=1-\frac{\left|{\mathrm{v}}_{\mathrm{t}}^{\mathrm{i}}-{\mathrm{v}}_{\mathrm{t}}^{\mathrm{r}}\right|}{\left|{\mathrm{v}}_{\mathrm{t}}^{\mathrm{i}}\right|+\left|{\mathrm{v}}_{\mathrm{t}}^{\mathrm{r}}\right|}$

It is observed that a high similarity result creates more chances that the corresponding laser cluster is our target object. For further improvement of the system, we updated the weight of the particles by selecting the clusters having the best similarity results as potential objects.

4. Experiment Details

We have checked and verified our approach by executing different experiments. The experiments were performed on a SCITOS G5 service robot (manufactured in Metralabs GmbH, Ilmenau, Germany), as shown in Figure 5. The robot has a 2D laser range finder (SICK S300), two circularly polarized antennas (RFMAX SS8688P), and a UHF RFID reader (Speedway Revolution R420) which has a sampling frequency of 2 Hz and has reading range up to 7 m (meter). All experiments are conducted with Dense Reader Mode 8 (DRM8) and the RFID reader channel is fixed at 920.625 MHz. The RFID antennas and tags are at same height of 1.2 m above the ground. The measuring range for the laser range finder is up to 29m, within the angle of 270° and resolution of 0.5°. In addition, the laser range finder works at a frequency of 20 Hz. A human, which is our target object, was equipped with four tags (Alien Squiggle RFID Wet Inlay from Alien Technology, San Jose, CA, USA) in the experiments.

As shown in Figure 5a, the SCITOS G5 robot already has all the required sensors and setup so it was convenient to use this robot instead of setting up a separate RFID system. We setup a rectangular area of 4 × 2m for experiments and setup the movement path as follows.

• TEST 1: Simple rectangular trajectory.

• TEST 2: Crossed rectangular trajectory.

The robot was placed at a distance of 1 m perpendicular to the test area, as shown in Figure 5. The human moved for 4 rounds at simple rectangular path and crossed rectangular path. Experiments were also conducted for both (clockwise and anticlockwise) directions of the concerned path. The human, equipped with an RFID tag array, stopped at some points but mostly moved with a velocity of approximately 0.4 m/s. In another set of experiments, multiple humans moved in the testing area and the effect of multiple persons was also verified for the proposed method. As the method used similarity results between laser cluster radial velocity and RFID phase-based radial velocity, it was observed that the best matching cluster was our target object.

4.1. Evaluation of Tracking Accuracy

Figure 6 and Figure 7 show that the estimated paths were almost consistent with true paths for the both simple rectangular trajectory and the crossed rectangular trajectory.

In Table 3, the average localization error represents rooted mean square error and it is defined as the Euclidean distance between the ground truth and the estimated position. It is observed that, for both TEST1 and TEST2, localization accuracy was significantly improved, as shown in Table 3. For every TEST, we performed different experiments and checked the results with different combinations of tags, starting from a single tag to four tags. Table 3 gives detailed results of different combinations of tags in different TESTs. Results show that using two tags is better than single tag and the usage of three tags is better than two tags. Four tags also give slightly better results than using three tags. It also strengthens the argument that using more than four tags may not be very useful because the “percentage of improvement in accuracy” reduces to an almost minimum level when the number of tags is increased from three to four. In a column of Table 3 named as “RFID Tag Combinations”, the letters a, b, c, and d are considered as RFID tag id attached at left hand, right hand, front, and back of the human, respectively.

4.2. Impact of Number of Moving Humans

We verified this proposed method for another possible scenario. There is a possibility of multiple humans moving around the target object so we checked the impact of a different number of moving humans in the testing area. It can be observed from the results of Table 4 that the number of moving humans has a small effect on the accuracy. In the first set of experiments, a human randomly moved in the test area at the same time as the target object was moving (as shown in Figure 5c) and the accuracy became slightly better (from 0.38m to 0.37m) as one human did not disturb the tag readings and the clustering results got better due to the increase in moving objects in the area.

In second set of experiments, when two persons moved randomly in the test area (as shown in Figure 5d) then the laser clustering got better but there were more chances that a human would block the RFID readings at some moments so localization accuracy slightly decreased from 0.38m to 0.39m.

4.3. Impact of Different Parameters

We checked the impact of different important parameters in this experiment. The impact of number of particle, K best matching clusters, and DBSCAN parameters were checked for the rectangle with simple rectangular path (TEST 1) in clockwise direction.

4.3.1. Impact of Number of Particles (N)

We checked the system for different number of particles. Since we use particle filtering for the tracking, the number of particles N directly affects the localization accuracy and the running time of the algorithm. The results are shown in Table 5. We used a CPU with Corei5-3337u (1.8GHz) with 4GB RAM for the processing. The other parameters of the experiments are set as follows: ε = 0.2, ζ = 2, K = 4, and ${\mathsf{\sigma }}_{\mathrm{d}}$ = 0.05.

It is observed that number of particles has a significant impact on localization accuracy. When N is less than 50, then the localization accuracy is not satisfactory and accuracy gets better when N is greater than 100. The best localization accuracy can be observed when the value of N is in between 200 and 500. Simultaneously, execution time is always smaller when the number of particles is reduced. We can see from Table 5, when value of N is set to 200 then it gives the best result in the minimum execution time of 12.82 milliseconds.

4.3.2. Impact of the Best Matching Clusters (K)

K and ${\sigma }_d$ are two important parameters to update particle filter. Table 6 shows the impact of the localization accuracy under this parameter. The remaining parameters are set as follows: ε = 0.2, ζ = 2, and N = 200.

It can be seen from Table 6, K and ${\sigma }_d$ have a significant effect on the localization accuracy. As there are no obstacles in the test area or in the way of the moving object, k>2 will give the best result. Translation coefficient ${\sigma }_d$ gives the best result of 0.38m when it is set to a value of 0.05 so k=4 and ${\sigma }_d$ =0.05 can be used to get the best localization accuracy.

4.3.3. Impact of Epsilon ε and MinPoints ζ

DBSCAN clustering is controlled by two parameters: Epsilon ε and the minimum number of points ζ. DBSCAN parameter setting depends upon the clustering requirements of the system. In this system, laser data is being clustered and it requires several experiments to be performed to test the effect of ε and ζ on the localization accuracy. The results are shown in Table 7. We set other parameters as follows: N = 200, K = 4, and ${\mathsf{\sigma }}_{\mathrm{d}}$ = 0.05. As can be seen from Table 7, a setting of ε = 0.2 and ζ = 2 gives the best localization result. We can see that a very small value of ε (e.g., 0.05) has a high impact on the localization accuracy and when its value is equal to or greater than 0.2, the impact is almost stable, provided that minPoints ζ is less than 4. It is also observed that a very high value of minPoints ζ (for example, 10) gives very bad results; on the other hand, a too-small value of minPoints ζ also increases the error, and the best results are at a value of 2.

5. Conclusions

In this paper, we proposed a different approach to localize a moving human equipped with an RFID tag array. We have calculated the phase-based radial velocity of RFID tags and compared it with laser-based radial velocity. Laser information was clustered through the DBSCAN algorithm. The best matching clusters were measured and fused with phase-based radial velocity in a particle filter. It was observed that for an area of 4 × 2m with a simple rectangular path, the tag array method gave a 25% increase in localization accuracy compared to a single tag approach. We plan to embed signal strength in this method and may also move the robot and object at the same time in upcoming work.

Author Contributions

S.Ur.R. analyzed the data and wrote the paper. R.L. proposed the idea and designed the experiments, G.L. and Y.F. conducted the experiments. H.Z. and A.Q. reviewed the paper.

Funding

This work is funded by the National Natural Science Foundation of China (No. 61601381 and 61701421), and partially by China’s 13th Five-Year Plan in the Development of Nuclear Energy under the grant number of 2016(1295)

Acknowledgments

We are thankful to the members of “Key Laboratory of Robot Technology used for Special Environment” in Sichuan province of China for facilitation and support to conduct experiments for this research.

Conflicts of Interest

The authors declare that there is no conflict of interest regarding the publication of this paper.

Abbreviations

RFID	Radio Frequency Identification
DBSCAN	Density Based Spatial Clustering of Applications with Noise
UWB	Ultra wideband
GPS	Global Positioning System
RSS	Received Signal Strength
UHF	Ultra High Frequency
TOA	Time of Arrival
TDOA	Time Difference of Arrival
AOA	Angle of Arrival
SAR	Synthetic Aperture Radar
BLE	Bluetooth Low Energy
IMU	Inertial Measurement Unit
MM	Magnetic Matching
PDR	Pedestrian Dead Reckoning
iGPS	indoor Global Positioning System
PF	Particle Filter
DRM	Dense Reader Mode

Figures and Tables

Figure 1. System flowchart giving a brief overview of the system.

Figure 2. Conceptual view of tagged human. Front and back view of a human equipped with a tag array and red rectangles represent radio-frequency identification (RFID) tags labeled as a, b, c and d. These red rectangles also express the body places where RFID tags are attached.

Figure 3. Raw Laser Data. It shows laser points readings during the experiment when multiple humans are moving in the test area.

Figure 4. Clustered Laser Data. It shows the laser points after applying DBSCAN (Density-Based Spatial Clustering of Applications with Noise).

Figure 5. Robot and experiment setup. (a)An example of a real time experiment; (b) conceptual view of the environment; (c) another human moving in the test area; (d) multiple humans moving in test area.

Figure 6. Comparison of ground truth and estimated track for 4× 2m with a simple rectangular trajectory.

Figure 7. Comparison of ground truth and estimated track for 4× 2m with a crossed rectangular trajectory.

Table 1

Comparison of related work.

Author/s	Approach	Accuracy (meters)	Remarks	Human Tracking
Shen [18]	AOA technique	0.13m	For static objects only	No
Ma [20]	TDOA	0.46m	Multiple reference tags are needed	No
Kaltiokallio [22]	RSS and combination of Kalman and Particle Filter	0.12–0.53m	Recursive and large amount of reference tags are required	Yes
Hsiao [23]	Real-time location system	0.3m	Multiple reference tags and multiple antennas are used	Yes
Wu [26]	Phase position model sensors	0.096m	Can work for static RFID tag only	No
Xiao [27]	Multipath propagation model	0.06m	For static items only and many antennas are needed	No
Liu [30]	Fusion of laser and RFID	0.2–0.65m	Specific environment and tracking of single tag	Yes
Wu [31]	Fuzzy reasoning algorithm and fusion of laser and RFID	0.09m	Works for robot localization	No
Wu [36]	Graph-based mapping with RFID and PDR	1.2m	Localization accuracy is limited.	Yes
Our Approach	Fusion of RFID tag array and laser	0.38m	No reference tags needed, model free and tracking of dynamic moving object	Yes

Table 2

Definition of mathematical symbols.

Mathematical Symbols	Definitions
$X_t$	Position of the object at time t
$g_t$	Measurement of laser at time t
$u_t$	Motion information of the object at time t
${\eta }_t$	Normalization factor at time t
$r_t^{tag}$	Measurement of RFID tag at time t
${\omega }_t^{\left[n\right]}$	Weight of particle n at time t
$X_t^{\left[n\right]}$	State of the particle at time t
$x_t^{\left[n\right]}, y_t^{\left[n\right]}$	Location of particle n at time t
N	Number of particles
K	The best matching K clusters
${\sigma }_t$	Gaussian noise at time t
$\tilde{l}$	Nearest cluster
$v_t^i$	Laser-based radial velocity of cluster i at time t
$v_t^r$	Phase-based radial velocity of the tag at time t
$C_t^i$	Laser cluster i at time t
${\sigma }_d$	Translation coefficient to the distance function
ε	Epsilon represents minimum distance between two laser points
ζ	MinPoints is minimum number of laser points to form a dense region
( ${\overline{x}}_t^i$ , ${\overline{y}}_t^i$ )	Center of cluster i at time t
${\phi }_t$	Signal phase at time t
$c$	Velocity of light
$f$	Signal frequency

Table 3

Evaluation of the average localization error (meter).

Total Tags	RFID Tag Combinations	Average Localization Error of TEST 1 (m)		Average Localization Error of TEST 2 (m)
		Clockwise	Anticlockwise	Clockwise	Anticlockwise
1	a	0.48	0.76	0.61	0.59
	b	0.86	0.50	0.57	0.63
	c	1.50	1.60	1.33	1.28
	d	1.75	2.11	1.84	1.17
2	a + b	0.41	0.45	0.53	0.56
	a + c	0.44	0.60	0.57	0.58
	a + d	0.45	0.62	0.59	0.60
	b + d	0.63	0.46	0.55	0.62
	b + c	0.60	0.45	0.54	0.61
	c + d	0.77	0.74	0.92	0.84
3	a + b + d	0.40	0.44	0.51	0.55
	a + b + c	0.39	0.43	0.50	0.54
	a + c + d	0.43	0.50	0.55	0.58
	b + c + d	0.47	0.45	0.53	0.61
4	a + b + c + d	0.38	0.41	0.49	0.52

Table 4

Impact of Number of Moving objects in test area.

Total Tags	Total Number of Moving Human	Average Localization Error of TEST 1 (m)		Average Localization Error of TEST 2 (m)
		Clockwise	Anticlockwise	Clockwise	Anticlockwise
4	2	0.37	0.40	0.47	0.50
4	3	0.39	0.44	0.51	0.54

Table 5

Effect of particle numbers N on localization accuracy and running time.

N	5	50	100	200	300	500	800	1000
Error(m)	4.39	0.41	0.39	0.38	0.39	0.38	0.40	0.39
Time(ms)	9.52	10.14	11.01	12.82	14.68	17.93	23.16	26.73

Table 6

Effect of K and ${\sigma }_d$ on localization accuracy.

		${\sigma }_d$	0.01	0.05	0.1	0.5
	Error (m)
K
1			1.26	0.48	0.86	7.01
2			0.41	0.41	0.41	0.47
3			0.39	0.39	0.39	0.44
4			0.40	0.38	0.40	0.45
7			0.41	0.39	0.40	0.45
10			0.41	0.39	0.43	0.47
13			0.41	0.40	0.39	0.45

Table 7

Effect of ε and ζ on localization accuracy.

		ζ	1	2	3	4	7	10
	Error (m)
ε
0.05			0.61	0.85	3.98	4.05	4.07	4.21
0.1			0.45	0.39	0.66	0.69	4.02	4.08
0.2			0.41	0.38	0.39	0.43	2.72	3.90
0.3			0.39	0.39	0.39	2.38	1.82	3.92
0.5			0.39	0.39	0.39	0.42	8.57	6.68