Tracking multiple dynamic objects by a combination of laser ranging and UHF RFID phase information

Radio Frequency Identification (RFID) achieves the identification of objects through electromagnetic waves and has wide applications in many areas. The system reports a list of tags in proximity without any knowledge about distance or bearing of the object. This paper presents an approach to integrate RFID phase and laser range information for the tracking of dynamic objects in an environment. The proposed system determines locations of objects by comparing the velocities estimated from two different systems. In particular, the laser range data is segmented into clusters using DBSCAN (Density-based spatial clustering of applications with noise). We compute the radial velocities of these clusters and compare them to the radial velocity estimated from RFID phase difference. The particle filtering is used to fuse the ambiguous phase measurements and to track moving objects with multiple hypothesis. The proposed approach uses the commercial off the shelf RFID devices and does not require the modelling of radio signal propagation. Experiments were conducted with a SCITOS G5 robot to verify the feasibility of the approach. The results showed that our approach can achieve a positioning accuracy of approx. 0.37 meters in a complex environment.


Introduction
Radio Frequency Identification (RFID) renovates the traditional ways of inventory and object identification [1] [2].Compared to the visual barcode, it provides a cost-effective solution for the automatic and contactless identification of objects.Particularly, the RFID working at the UHF band (Ultra high frequency, 860 to 960 MHz) has been widely used due to its long reading range.However, localization of these tags is not trivial, as the reader does not report any position information of the tag.
Previous researchers use proximity-based approach for the localization of the tag [3].Modern RFID readers report the signal strength and phase information, which are extensively examined for the positioning [4].A group of researchers approximates the distance with a distribution of the signal strength [3].To model the signal propagation, a number of factors has to be considered, such as different reader and tag models as well as various antenna gains.Also the environment introduces multipath propagation issue to the radio signal, for example water absorbs the signal and metal reflects the signal [5].This challenges the usage of received signal strength for localization.Phase is also a useful information to infer the distance, which is examined by another group of researchers [6].But phase is a 2π-periodic function, which means the phase repeats at distances with an interval of one-half the wavelength and cannot be used for positioning directly.Liu et al. [7] proposed BackPos to localize RFID tags using hyperbolic positioning technique based on the detected phase.Multiple antennas are required to resolve the phase ambiguity, resulting in increased cost overhead.Scherhäufl et al. [8] used a uniform linear array to handle the phase ambiguity.
Range-based sensors are widely used for object detection and localization.Some researchers also proposed to use laser range measurements for object recognition [9].To identity an object, a prior knowledge (i.e., model) about the object has to be obtained in advance.However, this process requires feature extraction and machine learning algorithm, which is time-consuming and computationally expensive.To improve the positioning performance, many researchers also focus on the incorporation of other sensory information.Liu et al. [10], [11] used RFID and laser to follow an object attached with an RFID tag.Fu et al. [12] achieved the positioning of a tagged-object by comparing the velocity similarity between two different systems, i.e., a laser range finder and an RFID reader.In this paper, we present an approach to combine phase and range measurements from RFID reader and laser range finder to track multiple dynamic objects in an environment.Figure 1 illustrates the objective of this paper.One application scenario of the proposed approach is human robot interaction at an exhibition or a conference.In these public events, different persons are wearing name badges affixed with RFID tags.Our approach offers an opportunity to effectively track the positions of persons even if they have similar visual appearance.With the location information, the robot could perform some friendly services such as offering some drinks, following a specific person.
One can determine the distance to object precisely by using a laser range finder.But distinguishing the tracking object from a large number of objects in the environment needs a prior knowledge about the target (i.e., shape).The RFID, on the other hand, provides the identity of the object, but without any location information.We achieve the tracking of an object by comparing the velocities estimated from a laser range finder and an RFID system.Rather than using the absolute RFID phase value, only phase difference is used.We employ the particle filtering to deal with the phase ambiguity issue.One may argue that tracking each tagged object with a particle filter is time consuming, as there might be hundred of objects affixed with RFID tags.But at a given time, the number of tags in the range is limited and a good accuracy can be achieved with a small number of particles as shown by our experimental results.The contributions of this paper are summarized as below:  We present an approach to track multiple dynamic objects by incorporating RFID phase and laser range measurements;  The usage of the particle filtering allows multiple hypothesis tracking and dealing with the phase ambiguity;  We thoroughly evaluated the performance of the proposed approach on a SCITOS G5 robot.

System overview
As illustrated in Figure 2, the proposed system makes use of the laser range and RFID phase information for multiple dynamic objects tracking.RFID provides a unique identity of the object in the view, but determining the locations of these tags is challenging, as neither distance nor bearing information is returned by the RFID system.Each tag is assumed to be affixed with one object for the purpose of recognition.Without any external knowledge, it is difficult to differentiate the tracking object (i.e., object with RFID tag) from a variety of objects in the environment.How to assign the object to the right RFID tag is regarded as data association problem.
The positions of these objects can be easily obtained via visual or range sensors.This paper uses DBSCAN (Density-Based Spatial Clustering of Applications with Noise) to segment the laser range data into various clusters.The tracking of object is achieved by comparing the velocity similarity between two different systems, i.e., a laser range finder and an RFID reader.The underline idea is that each moving object follows a unique trajectory.By examining the difference of these tracks (i.e., in this paper radial velocity), we are able to track the locations of different objects.We adopt a particle filter to track the multiple hypothesis.The algorithm allows to keep the tracking of multiple moving objects that have similar moving patterns.
As a preliminary result, we asked two users to move along the same rectangular path and showed the radial velocity between ground truth and estimation from RFID phase difference using our approach in Figure 3.As can be observed from this figure, the velocity variance between individuals is quite large, which can be used as feature to differentiate one user from another user.Another insight is that RFID radial velocity is quite consistent with the true radial velocity, which can be used as a good measurement of a user's movement.

Tracking with a Bayesian Framework
Given a series of measurements  : (including RFID phase measurement  : and laser range measurement  : , i.e.,  : ,  : ) and motion information  : up to time , our goal is to estimate the location of an object  at time .This problem can be formulated as the estimation of the probability density function   | : ,  : based on a history of sensor measurements and motion measurements.According to the Bayesian theory,   | : ,  : can be recursively updated as: where  guarantees that total probability sums up to one.  | ,  represents the motion model, which predicts the current state  given the previous state  and the motion information  at time .  | is the sensor model, which characterizes the likelihood of receiving a measurement  (i.e., RFID phase measurement  and laser measurement  ) given the state  .

Tracking with Particle Filtering
A number of implementations of Equation 1 can be found in the research community.The location of the object is represented by a set of N particles, i.e.,   .Each particle  consists of 2D position hypothesis and associated weight, i.e.,  ,  ,  .We perform the particle filtering with prediction and update based on the RFID and laser range measurements.To handle particle degeneracy problem, resampling is performed when the effective sample size [13] drops below .After this step, a new set of particles is generated based on their weights.

Laser clustering and radial velocity estimation
To accurately localize dynamic objects, the laser range measurement is segmented into various clusters.This step is achieved with DBSCAN, which is a density-based approach to group together the points that are close in space.Two parameters, namely ϵ and , are required to perform a DBSCAN algorithm [14]. represents a distance threshold to search for the neighbors of a point.If the distance between two points is smaller that , they will be assigned to the same cluster.To form a dense region, the minimum number of points is defined as .This step produces a set of clusters

𝐶 𝐶
, where  denotes the number of clusters.We use  , and  , to denote the 2D location of a cluster  .Figure 4 shows one clustering result using DBSCAN.
Next, we will estimate the radial velocity of a cluster based on the radial distance.For each cluster  , we find the nearest cluster  j at the previous timestamp  1 using: where  =  ,  , denotes the radial distance from cluster  to the origin.

RFID radial velocity estimation from phase difference
Phase information is available in most commercial RFID readers (i.e., Impinj R420 and ThingMagic M6).We represent the phase of tag as: where  represents the distance from the tag to the reader.Due to the two-way communication adopted in RFID systems, the total distance that the RFID signal travels is 2.The carrier wavelength is represented by λ, which is determined by reader frequency and electromagnetic wave speed (i.e., 3.0 10 m/s).For example, the wavelength λ at a frequency of 920.5MHz is 0.326 meters.The RFID transmitter, tag, and receiver circuit additionally introduce a certain amount of phase rotation, which are described as  ,  , and  respectively.The phase value falls between 0 and 2.
,  , and  are assumed to be fixed for a tag and RFID reading frequency.Let  and  denote the phase values at t and  1 respectively, the radial distance traveled by the tag during time ∆ can be expressed as: The above equation shows the radial distance can be computed by the phase difference plus an additional 2 phase, where  is an unknown integer.Assuming that the phases are measured in a very short period of time and the movement of the object during this time is less than a wavelength (i.e., λ), we can treat  as zero.Therefore, the radial velocity of the tag  can be computed as: The equation above normalizes the phase difference to between  and .A positive radial velocity indicates the object is moving away from the reader and a negative radial velocity tells the object is moving close to the reader.Please note that, to obtain the phase difference, one has to measure the phase at the same channel (i.e., same frequency) and the same antenna.
In addition, the decoding technique used in most commercial readers (including Impinj and ThingMagic) introduces an ambiguity of  in phase.In another word, the measured phase can be either the true phase or the true phase plus  rotation, which is known as phase flip issue.We show the estimated radial velocity in Figure 5 using the raw phase difference.As can be observed from this figure, the radial velocity can be quite different from the ground truth due to the random  phase flip issue.This ambiguity strongly limits the use of phase in many applications.We notice that it is very rare that the true phase difference is close to  at two continuous phase measurements in a short period of time.Therefore, we ignore a phase measurement if |∆ | exceeds a threshold .

Motion model
We predict the state of a particle  according to the previous state  and a motion model   | ,  .The prediction is performed as the following: where we denote  0,  as a Gaussian distribution with zero mean.The standard deviation is represented by .A suitable  is expected at the prediction stage to capture the movement of a user.We evaluate the positioning accuracy under the impact of  in the experimental section.

Sensor model to update particle weights
The measurement  comprises two sources of information, i.e., the RFID phase  and the laser range measurement  .According to the conditional probability, we further decompose   | into: where  is a parameter that allows the particles to focus on the region of laser cluster  .To model the likelihood of receiving a phase  given the location of the object is not straightforward, since the phase is a 2 -periodic function to the distance and is determined by many factors, for example tag and receiver rotation as shown in Equation 4. Instead of using the absolute phase value, we use the relative phase change to model   | ,  .We compute the radial phase velocity and compare it with the true CTIS-2023 Journal of Physics: Conference Series 2595 (2023) 012007 IOP Publishing doi:10.1088/1742-6596/2595/1/0120077 radial velocity, which is approximated by the radial velocity of the cluster.We use the following formula to compute the similarity between two velocities, namely the radial velocity  of a cluster  and RFID phase radial velocity  : The idea behind this is that: A number of static or moving objects exist in an environment.Given the radial velocity of the tag, the cluster with the closest velocity is considered to be the tracking object.

Experimental details
We verified the performance of the proposed approach on a SCITOS G5 robot from Metralabs, Germany.The robot is equipped with a SICK S300 laser range finder, as shown in Figure 1.The laser range finder offers a maximum range up to 29 meters.An RFID reader (Impinj Speedway Revolution R420) is installed on the robot for phase sensing.We attached the reader with two RFID antennas (Lairs Technologies, 9 dBi antenna gain).The antennas are placed with a height of approx.1.2 meters with a spanning angle of ±45° to the facing direction of the robot.We set the frequency of the reader to 920.5 MHz and use the maximum transmitting power of 30dBm.The radial velocity estimated from RFID phase is passed through an exponential smoothing filter with a smoothing factor of 0.5.
Three persons were asked to walk along a rectangular path with a size of 4 2 meters at a normal walking speed.We attached each user with an Alien Squiggle RFID tag.We recorded 2550 RFID measurements in 130 seconds.Each user walked a total distance of approx.36 meters.The laser range finder reports four scans every second.We placed several landmarks with known locations along the walking path as the ground truth.The user was asked to press a button on a phone to record the landmark position when he walked through the landmark.We measure the positioning error by the Euclidean distance between the estimation and the ground truth.The mean positioning error is shown as the average of the positioning error among three users during the whole walking path.

Impact of σ and 𝜎 on the positioning accuracy
In our approach,  and  control the particle prediction and update, respectively.In the first series of experiments, we examined the impact of  and  on the positioning accuracy.The number of particles N is set to be 1000.We fix  = 0.5 (in meters) and  2 in DBSCAN.Table 1  As can be seen from this table,  1.0 gives the best positioning accuracy.This setting coincides with the actual human walking velocity (1.0 m/s).A too large or too small  leads to a reduction of the positioning accuracy.The reason for this can be explained as follows: A too large  introduces a high degree of uncertainty to the motion model and results in a divergence of the particle filtering and a decrease of the positioning accuracy.A small  is usually not fast enough to capture the motion of a user, which also leads to a large positioning error.In addition, we observed that  0.3 produces the best positioning accuracy.A too large or too small  results an increase of the positioning error.A large  is not able to fully utilize the advantage of the laser range information.A small  gives too much trust on the laser range information, which might not be correct due to the uncertainty of range measurements and the occlusion from other users.
The estimated tracks of users are shown in Figure 6(a) and the positioning errors at different times are shown in Figure 6(b).As can be observed from this figure, we obtained a very large positioning error at the beginning (i.e., first 20 seconds) of the estimation, as the particle filtering keeps tracking of object with multiple hypothesis and is not able to determine the right object to be tracked; after the integration of the RFID measurements, we are able to precisely localize all users at the same time (see the decrease of the positioning error after  20).

Impact of particles N on the positioning accuracy
In the next experiments, we examined the positioning accuracy and the running time with respect to the number of particles N. We fixed  1.0 and  0.3.We run the algorithm on an Intel Core i5-4200M 2.50GHz CPU with a RAM of 4GB.The positioning result is shown in Table 2. From this table, we can see that the positioning accuracy decreases with the reduce of number for particles.However, with particle numbers larger than 500, the positioning result stays stable.From Table 2, we can also see that performing the algorithm with one sensor update (including laser clustering, particles prediction, and particles update) with  1000 only consumes 19.62 ms.For the RFID reader used in our experiment, interrogating one tag using our reader takes approximately 50 ms.Thus, our algorithm can be applied in real time applications.

Impact of 𝜖 and 𝑚𝑖𝑛𝑃 on the positioning accuracy
The clustering of the laser range measurements is controlled by two parameters, i.e., distance threshold  (in meters) and minimum number of points .Next, we examined the positioning accuracy by varying  and  in DBSCAN.The number of particles is fixed to be  1000.We set  and  0.3 .Table 3 shows a comparison of the positioning results.As can be observed from this table, the best positioning result is achieved with a setting of  0.5 and  2. The choice of  has a high impact on the positioning result.A too small or too large  deteriorates the positioning results.With a large  1.5, several objects are merged into one cluster, which will lead to a decrease of the positioning accuracy.A small  0.05 might segment an object into different clusters, and leads to an increase to the positioning error.As can be also observed from Table 3, a small  slightly decreases the positioning accuracy.A large  is not suitable for clustering small objects (like humans) and leads to a deterioration of the overall positioning accuracy.

Impact of phase difference threshold ϑ on the positioning accuracy
We performed experiments to examine the tracking performance under phase difference threshold ϑ.The number of particles N is set to be 1000.We choose  1.0 and  0.3.The positioning result is shown in Table 4.As can be observed from this table,  provides the best positioning result.A too large and too small ϑ results in a poor positioning accuracy.A small ϑ will filter out the valid phase measurements and results in a low positioning accuracy. 0 ignores all phase measurements, which is equivalent to a system that only uses laser scan for tracking. 0 gives a positioning accuracy of 2.25 m, which is much worse than our proposed approach, since we cannot distinguish the tracking object from other objects by only using laser range data.A too large  also gives a bad positioning result, as in this case too much ambiguous phase measurements are integrated, which leads to a poor approximation of the sensor model and a decrease of the positioning accuracy.

Conclusion and future work
This paper proposed an approach to fuse RFID phase and laser range information for the tracking of multiple dynamic objects.RFID provides a cost-effective solution for identification of objects affixed with tags.But it is difficult to determine the locations of these tags.The laser range finder provides a precise distance information, but is not able to know the identity of the object.A combination of these two sensors allows to generate an accurate and new tracking system.For this purpose, we segmented the raw laser range measurements into different clusters with DBSCAN.The radial velocities of these clusters are compared to the radial velocity computed from the RFID phase difference.We adopt a particle filter to combine these two measurements to track the location of an object continuously.Our approach annotates the cluster with the correct RFID tag by exploiting the moving pattern of a person.We validated the proposed approach on a robot that equipped with an Impinj RFID reader and a SICK laser range finder.Our experimental results show that we can achieve the positioning of a group of users with an accuracy of 0.37 meters.One of the future work is to design a controller that enables the robot to follow a moving human.

Figure 1 .
Figure 1.Illustration of the objective: our goal is to track the locations of multiple moving objects (i.e., three persons as an example) affixed with RFID tags by the integration of RFID and laser.

Figure 3 .
Figure 3.A comparison of the radial velocity estimated from RFID phase difference and the ground truth.

Figure 5 .
Figure 5. Ground truth and the estimated RFID radial velocity based on RFID phase difference without considering the  phase flip issue.

Figure 6 .
Figure 6.Evaluation of positioning accuracy.(a) Estimated track and the ground truth; (b) Positioning accuracy of different users at different timestamps.
| specifies the likelihood of observing a laser cluster  given the location of the object  .Assume the laser beam cannot go through the object, therefore we model   | as:

Table 1 .
lists a comparison of the positioning accuracy under different settings of  and  .Average positioning accuracy in meters with respect to  and   .

Table 2 .
Positioning accuracy (in meters) and the execution time (in milliseconds) of the algorithm with respect to different number of particles N.

Table 3 .
A comparison of the average tracking accuracy (in meters) with respect to  (in meters) and  in DBSCAN.

Table 4 .
Tracking accuracy (in meters) with respect to phase difference threshold .