Measurement-Based LoS / NLoS Channel Modeling for Hot-Spot Urban Scenarios in UMTS Networks

A measurement campaign is introduced for modeling radio channels with either line-of-sight (LoS) or non-line-of-sight (NLoS) connection between user equipment (UE) and NodeB (NB) in an operating universal mobile telecommunications system. A spacealternating generalized expectation-maximization (SAGE) algorithm is applied to estimate the delays and the complex attenuations of multipath components from the obtained channel impulse responses. Based on a novel LoS detection method of multipath parameter estimates, channels are classified into LoS and NLoS categories. Deterministic models which are named “channel maps” and fading statisticalmodels have been constructed for LoS andNLoS, respectively. In addition, statistics of new parameters, such as the distance between the NB and the UE in LoS/NLoS scenarios, the life-distance of LoS channel, the LoS existence probability per location and per NB, the power variation at LoS to NLoS transition and vice versa, and the transition duration, are extracted.These models are applicable for designing and performance evaluation of transmission techniques or systems used by distinguishing the LoS and NLoS channels.


Introduction
Distinctive characteristics of signals coming from a line-ofsight (LoS) or non-line-of-sight (NLoS) path in a wireless environment raised great concerns in mobile radio system design.The probability of whether the LoS path exists in the channel or not depends on various factors, for example, terrain features, building density and locations of users.Statistical models have been employed by the 3GPP spatial channel model (SCM) [1] and the WINNER II SCM-enhanced [2] to describe the LoS existence probability.Recently, characterization of LoS/NLoS channel becomes highly demanded in some applications, such as massive multiple-input-multiple-output (MIMO), mobile back-hauling, user-specific beamforming, and other adaptive techniques used in future heterogeneous networks [3][4][5][6].Newly introduced multipath access scheme-nonorthogonal multiple access (NOMA)for future radio access also relies on the distinction of LoS and NLoS channels [7][8][9].Additionally, some ad-hoc wireless communication scenarios and the implementation of some new technologies also prefer the LoS channel, for example, dedicated short range communication (DSRC) system for device-to-device, vehicle-to-vehicle, and machine-tomachine communications [10][11][12][13][14][15].Furthermore, the LoS existence has been considered in localization applications, such as the distance estimation between the base station (BS) and the mobile station in [16].Modeling realistic characteristics of LoS/NLoS channels based on measurements are important for designing and optimizing these techniques or systems.
A problem arising when characterizing the LoS/NLoS channels based on measurements is how to categorize a channel as LoS or NLoS.Detection methods addressed in the literature can be split into two groups, that is, the range-measurement-based and channel-impulse-response-(CIR-) based methods.The former method, such as those introduced in

Models of channel parameters
Figure 1: A diagram of channel measurement, data analysis, and modeling procedure adopted for LoS/NLoS channel characterization.[16][17][18], was adopted for LoS detection in the early stage.The latter approach is mainly based on the shape or statistical properties of the CIRs, for example, the confidence matrix [19], the delay spread [19], the power delay profile (PDP) [20], and the narrowband -factor [21,22].In our contribution, the LoS detection is performed based on the power difference among estimated multipath components, which has been proven to be effective through measurement campaigns.
Some LoS/NLoS propagation models have been reported in the literature.According to measurement environments, these models characterize channels in either indoor or outdoor scenarios.In indoor cases, the distribution of the channel gains for the LoS/NLoS scenarios is presented [23,24].Coherence bandwidth and delay spread in LoS/NLoS scenarios have been studied at 5 GHz [25,26] and 60 GHz [27].In [28,29], statistical models characterizing the shadowing in LoS and NLoS scenarios were extracted from indoor measurements.In outdoor scenarios, path loss models have been attracting more attention [30][31][32][33][34]. Furthermore, statistics of the LoS coverage rate in urban areas for broadband wireless access systems and the LoS existence probability versus elevations of departure of downlink channels have been investigated in [35,36], respectively.
Experiment-based stochastic channel modeling relies on measurement campaigns to be conducted extensively in the environments of interest.Many conventional models have been extracted based on dedicated channel sounding activities.However, due to practical concerns, such as power constraints, poor mobility of the sounding equipment, and limited number of measurement routes, the channel observations do not possess sufficient randomness to characterize the channels ergodically, leading to limited applicability of the resultant models.As an alternative, using in-service public wireless communication systems or networks to measure channels in large areas can be adopted.In [37], a public base station was used as a transmitter, and the received signal strength was measured in the whole coverage of the base station.The advantage of channel measurements in in-service networks is that the channel characteristics observed are identical with those experienced by the user equipment (UE); thus the channel statistics extracted are more realistic.Compared with hundreds of MHz bandwidth usually adopted in the dedicated wideband channel measurement campaigns, the measurement based on the conventional 2G public system usually utilizes hundreds of KHz bandwidth, which facilitates the study of narrowband channel properties, such as the fading-related parameters [38] or performancerelated parameters, for example, the out-of-channel interference [39].Along with fast developing commercial systems, such as LTE-Advanced, wideband characteristics of channels can be measured based on the in-service networks.Channel modeling based on these data began to attract more attention recently [36,40].
In this work, a receiver was constructed to collect the downlink signals at a hot-spot area of Shanghai in a commercial in-service UMTS network.The impulse responses of the channels are calculated, and a high-resolution estimation algorithm derived based on the SAGE scheme is applied to extract multipath parameters.After appropriately detecting the LoS and NLoS channels by using the novel LoS detection, various aspects of LoS/NLoS channels are investigated and the corresponding models are established based on the extensive observations.They include the deterministic channel maps, which are the geographic map overlaid with spots color-coded by channel characteristics, and statistical models characterizing the channel gain, shadowing, and multipath fading.Furthermore, the statistics of some newly defined channel parameters are studied, which include (i) the distance between NBs and the user equipment in LoS and NLoS conditions, (ii) the life-distance of LoS and NLoS channels, (iii) the LoS existence probability per carrier for a UE, (iv) the LoS existence probability per carrier for a NB, and (v) power variation at LoS-NLoS transition and the transition duration.These models are useful for the design and performance evaluation of communication techniques adaptive to the LoS and NLoS scenarios.
The rest of this paper is organized as follows.Section 2 introduces the measurement equipment, environments, and the specifications applied for data acquisition in an in-service UMTS network.The main steps of processing the measurement data to obtain the estimates of channel impulse responses are also described.In Section 3, the LoS/NLoS detection method is introduced.Section 4 elaborates details of the models established based on the LoS/NLoS channel characteristics obtained.Finally conclusive remarks are addressed in Section 5.

Measurement Equipment, Campaigns, Setup, and Postprocessing
Figure 1 illustrates the diagram of the data acquisition, processing, and modeling procedure adopted for characterizing LoS/NLoS channels based on the signals of a UMTS network.The equipment used to acquire the signals consists of the following components: a Universal Software-Defined Radio Peripheral (USRP) device of type N210 [41] controlled by the GNU radio software in a computer and a global positioning system (GPS) module that reports the longitude and latitude of the measurement location to the computer through a serial port, the storage disk, an omnidirectional antenna working in 2-3 GHz UMTS frequency band, and a lead-acid battery.All this equipment was loaded in a trolley moving along predefined routes during the field measurements.
The measurement campaign was performed in the pedestrian zone along Nanjing Road, Shanghai, China.This area is a typical hot-spot of the local UMTS network where complex streets and densely distributed skyscrapers jointly create a city canyon environment.Multiple macro-, micro-, or pico-cellular UMTS stations were deployed in the area.Figure 2 shows a photograph taken on the Nanjing Road during the measurements.Measurements along the streets with different styles of corners provide a large amount of channel observations which are sufficient statistically for modeling the realistic LoS/NLoS channels, including the channel evolution characteristics during the LoS-to-NLoS transitions and vice versa.
During the measurements, the receiver moves at a walking speed of about 1.2 m/s.Data acquisition was triggered when the receiver departed from its previous location by 5 meters, and the duration of collecting the data at each location is 1 second.The carrier frequency was 2.1376 GHz and the receiving bandwidth was 20 MHz.The scrambling code existing in more than 5 consecutive frames is considered to be a valid code.It was found later that the data collected in the whole campaign was from 78 NBs, each with a distinctive combination of a scrambling code and a physical carrier.
The raw data collected during the measurements was post-processed in the following five steps.
Step 1 (filtering).The measurement results show that three UMTS carriers were used with central frequencies of 2.1326 GHz, 2.1376 GHz, and 2.1426 GHz, respectively.The raw data received with 20 MHz effective bandwidth at the center frequency of 2.1376 GHz is first filtered to obtain three baseband signals   (),  = 1, 2, 3, for these carriers individually.
Step 2 (synchronization).The beginnings of time-slots in the downlink synchronization channels are detected by correlating the baseband signals with the primary synchronization code (PSC).Multiple NBs are found by checking the dominant peaks in the PDPs of the obtained CIRs.In addition, the estimates of the starting time  0 of individual data frame in the common pilot channels (CPICHs) and the index  of estimating scrambling code group are obtained by solving the maximization problem: where    (),  ∈ [1, 2, . . ., 64], consists of 15 secondary synchronization codes permuted according to the th predefined sequence,  = 10 ms is the duration of a CPICH frame [42], and (⋅) * denotes the complex conjugate of given argument.
Step 3 (scrambling code detection).The detection of scrambling code is necessary in our case due to the fact that the information that how the scrambling codes are allocated in the carriers is unknown.The index  of the scrambling code used to modulate the CPICH is detected by where the vector  î = ( î − 1) ⋅ 8 + [1, 2, . . ., 8] contains the indices of 8 scrambling codes in the îth group and   () denotes the th scrambling code.Notice that, for some channels, although the index of scrambling code group can be detected in Step 2, none of the scrambling codes within the group exhibits significantly higher likelihood of being the correct code than the others.This is probably due to the fact that the frequency offset between the NBs and USRP is nonnegligible, resulting in low SNR or signal to interference and noise power ratio (SINR) in the CPICHs.In such cases, the data acquired at this location will not be considered for modeling.
Step 4 (CIR extraction).For the channels identified with the ĵth scrambling code at carrier , the impulse response ℎ , () of the channels observed in the th CPICH frame is calculated as Step 5 (parameter estimation).It can be shown that the channel coherence time is beyond the duration of one frame even in the worst case where the people in the environment are moving at the largest speed of 3 m/s.Thus, the channel observations obtained in one data frame are considered to be coherent and applicable for multipath extraction.A SAGE algorithm is derived and used to estimate the delays and complex attenuations of multiple specular propagation paths from the CIRs observed in individual data frames.Readers may refer to [43] for the derivation of the SAGE algorithm and the steps for implementation.Based on a practical principle that the paths estimated with power 20 dB lower than the maximum are negligible, we found that 10 paths per CIR are sufficient for reconstructing the dominant portion of the CIRs in most cases.Furthermore, observations not reported here also show that the parameter estimates usually converge within 30 iterations. Figure 3 depicts the comparison between the original PDP for the channel observed in a CPICH frame and that calculated from the reconstructed signals that are computed based on the estimated 10 specular paths.The stem plot for the powers of the estimated paths with the asteroid marker is also illustrated in Figure 3.It is obvious that the original PDP is well approximated by the reconstructed PDP, indicating that the path components estimated by using the SAGE algorithm can represent the channel impulse response accurately.

LoS/NLoS Detection
The LoS detection method considered here is based on the estimated multipath parameters.It is a common belief that the LoS path arrives at the UE with the shortest delay.However, due to the limited delay resolution of the communication signals, a LoS path may be combined with NLoS paths which have slightly larger delays than the former.Thus, a multipath cluster can be identified in the vicinity of the true LoS path.This is particularly the case in urban environments where the local clutter around the NB may generate a large amount of diffraction or reflection paths with delays close to the LoS path.Considering these situations, the LoS detection method adopted here is designed for two cases.In the first case where the first-arrival path has the largest power among all estimated paths, a clear decision that the channel is a LoS channel is made without further analysis.In the second case where the first-arrival path does not have the largest power, the deviation  between the powers of the strongest path and this first-arrival path is calculated and used to determine the channel scenarios based on a hypothesis testing.A predefined threshold  th is applied in the testing, where the channel is LoS for  >  th and NLoS for  ≤  th , respectively.The suitable value of  th was identified through the measurements collected at a corner on Nanjing Road, where the receiver moved from LoS to NLoS scenarios.Figure 4 depicts the empirical probability density functions (pdfs) of  observed in the true LoS and NLoS cases.It can be observed that the two pdfs interset to overlap around  = 9-12 dB.Our calculation shows that, by setting  th = 10 dB, the accumulative probability for detection error in this case can be kept less than 0.1.Thus, practically,  th = 10 dB is chosen for the hypothesis testing of a channel being LoS or NLoS in the environment considered here.
It is worth mentioning that the measurement environments were time-variant with people walking around, which can significantly influence the accuracy of LoS detection based on a signal observation.To improve the detection accuracy, it is necessary to select the decision in majority when multiple decisions are available.In the case considered here, the decision with higher occurrence frequency among 49 CPICH frames per location is selected.In most locations, more than 35 frames were observed to provide common LoS detection result.
By using the aforementioned detection method, LoS and NLoS channels observed with valid scrambling codes are identified for all measurement locations.Figures 5(a locations for the channel with the largest power among all valid channels identified.It can be observed that more LoS channels are found than the NLoS channels at the left end of the Nanjing Road.This is reasonable as that area is an open cross with two main roads intersect.Furthermore, NLoS channels are widely observed along the Nanjing Road, a typical street canyon environment where the diffraction and the reflections from skyscrapers are the major propagation mechanisms.

LoS/NLoS Channel Models
In   scenario, and the standard deviation of the channel gain in the LoS scenario is 4.65 dB less than that in the NLoS scenario.These statistical differences indicate that the existence of the LoS path in the channel can not only bring 6 dB gain in average for the received signal but also reduce the variation of the channel gain significantly.
To calculate the gain due to the shadowing   which attributes to the large structures around the receiver, it is necessary to subtract the path loss   caused by free-space propagation and the terrain features.In our case, since the exact locations of the NBs are unknown, the path loss may not be calculated analytically by using existing models.As International Journal of Antennas and Propagation  an alternative, the empirical path loss is computed in the following manner.For any location, say , channel gains observed in the surrounding locations within 50-m radius are averaged to obtain the estimate of the path losses  ,LoS and  ,NLoS at  for LoS and NLoS scenarios, respectively.The shadowing of a LoS channel  ,LoS at the location  is then calculated by taking the mean of  LoS, −  ,LoS , where  LoS, is the channel gain observed in a LoS scenario at any location less than 10-m deviation from .Similar calculations are applied to calculate the shadowing of a NLoS channel  ,NLoS .Figure 7 depicts the empirical cdfs of the shadowing in the LoS and NLoS scenarios, respectively.The cdfs of - distributions which were found to fit the best with the empirical graphs among multiple candidates are also illustrated in Figure 7.The parameters of the pdfs of the adopted distributions are reported in Table 1.It can be observed from Figure 7 that the mean of the gain due to shadowing both in LoS and NLoS scenarios are around 1 in linear scale, which indicates that the path loss has been correctly removed in the calculation.The spread of the shadowing is larger in the NLoS channels than that in the LoS channels.This observation shows that more complex environments exist around the receiver in the NLoS scenarios, which create larger variation of the shadowing than in the LoS scenarios.
Multipath fading in a channel can be calculated by subtracting the path loss and shadowing from the over channel gain.Figure 8 depicts the empirical cdfs of channel gain attributed to the multipath fading in the LoS and NLoS scenarios, respectively.The - distributions which were shown by the K-S test to fit the best to the empirical graphs are also depicted in Figure 8.It can be observed that the multipath fading magnitudes in the linear scaling are distributed within the interval of [0, 3] with mean values equal to 1 for both scenarios.Furthermore, the cdf is more concentrated in the LoS than in the NLoS scenarios, indicating that the variation of the multipath fading is more severe in NLoS scenarios.Table 1 Empirical data for LoS Empirical data for NLoS - cdf fitted to LoS - cdf fitted to NLoS reports the parameters of the - cdfs that fit well with the empirical data.

Distance between the NB and the UE.
As the locations of the NBs are unknown in our case, it is difficult to establish the path loss models for both LoS and NLoS scenarios.However, the distributions of the path loss, which can be calculated based on a method that will be introduced later, are applicable to deducing the statistics of the distances between NBs and UE in either LoS or NLoS cases.Such information is considered to be useful for channel simulations and designing the transmission power for NBs and UE with respect to the LoS and NLoS scenarios.
Here   can be considered to be deterministic,   depends on the direct distance  between the NB and the UE, and   ,   are uncorrelated random variables.In the case considered here, the unknown  can be viewed as a random variable, and consequently   , as a deterministic function of , is a random variable as well.The pdf of   is the convolution of the pdfs of −  ,   , and   .Under the condition that the empirical pdfs of   and   are known, we can calculate the pdf of   by the deconvolution operation.Figure 9 depicts the estimated empirical pdf of the   in LoS and NLoS scenarios.It is worth mentioning that, since   is unknown, the absolute value of the abscissa of Figure 9 may be incorrect.It can be observed that the mean path loss is 7 dB less in LoS than in NLoS scenarios and the spread of the path loss is larger in NLoS than in LoS scenarios.
Since the path loss is a deterministic function of the distance  between the NB and the UE, we may deduce the distribution of  from the estimated pdf of   .For simplicity, the following path loss model [44,Section 2.6] is used: International Journal of Antennas and Propagation where the path loss exponent  should be selected within [2.7, 3.5] and the reference distance  0 is between 10 and 100 m for urban micro-cellular environments.In our case,  = 3 and  0 = 50 are selected, which can be changed if more accurate path loss model is available for the environment considered.By assuming that the transmit power   = 0 dB, the pdf of  is obtained from the empirical pdf of   by replacing   with (4). Figure 10 illustrates the resultant pdfs of direct distance from the NB to UE in LoS and NLoS scenarios, respectively.It can be observed from Figure 10 that the LoS users are maximum 2.5 km from the NB.The most of the LoS users are located at 200 m from the NB.The NLoS users are widely spread within the range from 0 to 4 km, and their majority is around 300 m from the NBs.The nonzero probability observed for large range up to 4 km indicates that this may be due to the existence of macro-cellular NBs.These results clearly demonstrate that the LoS users are located closer to the NB than the NLoS users.In other words, the with small coverage areas tend to serve more LoS users than the NBs with larger coverage ranges in the Nanjing Road environments.

Life-Distance of LoS Channel.
To implement, for example, the user-specific beamforming technique, a UE is usually expected to experience LoS scenarios in a long distance.Thus, it is important to investigate the statistics of the so-called "lifedistance of the LoS/NLoS channel"   within which a UE experiences the LoS channel constantly.Based on the location information of the measured spots, the measurement routes are accurately reproduced.In the case where the LoS or NLoS channel appears in a certain number of consecutive locations,   can be calculated as the displacement from the first to the last location.Figure 11 illustrates an example of the fragments where LoS or NLoS channels were constantly observed for the NB with the scrambling code number 188.In this example, the maximum of the life-distance of LoS channel is calculated as  25 meters which was found along the spacious street.Notice that, due to the limited GPS resolution and the fixed distance of 5 m between adjacent measured spots, the minimum lifedistance of LoS or NLoS channel is lower-bounded by 5 m. Figure 12 depicts the cdf of the life-distance of LoS and NLoS channels calculated based on the observations of totally 148 fragments for LoS and 298 fragments for the NLoS channels.It can be observed from Figure 12 that the longest LoS and NLoS life-distance can be up to 28.1 m and 160 m, respectively.Furthermore, nearly 50% of the LoS and NLoS life-distances observed are 5 meters, and 20% of the LoS and NLoS life-distances are within the range of 5 to 10 meters.The majority of the life-distances observed are below 10 meters which indicates that the LoS and NLoS scenarios transit frequently when a UE walks on the city canyon environments as in the case of Nanjing Road.that the UE may receive signals from more than one NB, the "LoS existence probability for a UE"  LoS/UE is defined as the percentage of the NBs with the fact that the UE can set up a LoS channel among all valid NBs.In the measurement data, the case of receiving two scrambling codes per carrier is the majority.The  LoS/UE for the two-scrambling-code cases is investigated based on available observations.Figure 13 depicts the empirical percentage of  LoS/UE for three carriers.It can be observed that the distributions of  LoS/UE are similar for all three carriers, which is probability due to the fact that every NB in Nanjing Road makes use of three carriers.Furthermore, the case of no LoS path existed to the two NBs has the probability larger than 0.4.The opposite situation that all NBs have LoS connections to UE has the probability around 0.15.It is easy to deduce that the case where the UE has at least one LoS connection to the surrounding NBs can happen with probability 0.6.These results may be used to provide probabilistic assumptions for designing the cooperative multipoint (CoMP) techniques when the coexisting multilinks involve both LoS and NLoS channels.
For individual NBs, it is important to know the distribution of a specific percentage of UE that are categorized as LoS users among all UE served by the NB.The "LoS existence probability per NB"  LoS/NB is defined as the ratio of number of LoS UE to the number of all UE in the coverage of a NB. Figure 14 illustrates the empirical percentages of LoS existence probability per NB per carrier, calculated based on the observations in total 78 NBs.It can be observed from Figure 14 that, for three carriers, the situation of all UE in a NB experiencing LoS channel is zero, the situation in which 20% of the UE in a NB are LoS users occurs with 0.4 probability, and the case with 80% of the UE being LoS users happens with probability close to 0.2.This result indicates that, on the Nanjing Road, the existing NBs can be categorized into two groups, one with the most users being LoS users and the other with the most users being NLoS users.This categorizing scheme probability is due to the fact that the most NBs deployed on Nanjing Road are either micro-or pico-cellular NBs.Depending on the type of the NB, distinctive distribution of "LoS existence probability per NB" may be caused.
In addition, an interesting phenomenon observed from Figure 14 is that, for the carriers 2 and 3, higher  LoS/NB is observed with larger probability than in the case of carrier 1.This is probably due to the reason that the carriers 2 and 3 might be more widely used in the pico-cellular NBs than the carrier 1.Since more LoS users exist in the pico-cellular NBs than other types of NBs [45], the carriers 2 and 3 exhibit higher probability than the carrier 1 for larger  LoS/NB .power than the LoS scenarios.This phenomenon may be due to the fact that, for some locations such as the corners of streets, when a UE turns its moving direction significantly, the LoS-to-NLoS transition occurs, and the constellation of multipath propagation may include new NLoS paths which add more power to the channel than the power lost due to the absence of the LoS path.This is different with the obstruction cases where the overall propagation constellation is retained with only the LoS path absent due to blockage.

LoS-to-NLoS Transition Duration.
The LoS-to-NLoS transition duration  is defined as the time spent on the period from the starting of transition to the moment that the NLoS channel begins to stabilize.Here the channel is considered stabilized if the LoS or NLoS status is maintained for no less than 10 frames.The observations show that all LoS-to-NLoS transitions can be categorized into four types depending on the times of back-and-forth between LoS and NLoS scenarios.Figure 16 illustrates examples for the four transition types.Type (i) is referred to as the case where a LoS channel changes to a NLoS channel which stabilizes immediately.For such cases,  is regarded to be 1 frame.Type (ii) denotes the situation that, after the transition starts, the NLoS channel changes back to the LoS scenario once and then transits again to the NLoS scenario.For this case,  is considered to be 3 frames.Type (iii) is similar with type (ii) but the difference is that, during the transition, the NLoS scenario lasts for two frames before transiting back to the LoS.Type (iv) is similar with type (iii) except for the fact that the LoS channel remains for two frames before transiting to the NLoS scenario in the second time.For the latter two types,  is considered to be 4 and 5 frames, respectively.Figure 17 depicts the empirical probability of  for the four types considered.It can be observed that the type I transition occurs with the highest probability of 0.9 among all types.Only a small amount of transitions undergoes the fluctuations in types (ii) to (iv) which cost a longer time for channel being stabilized.These results show that, for the walking speed in a city canyon environment, most of the LoS-to-NLoS transitions are accomplished rapidly in one frame, that is, 10 ms.Longer transitions lasting for more than one frame can be neglected when designing systems for their low occurrence probabilities.

Conclusions
A channel measurement campaign has been conducted aiming at characterizing LoS/NLoS channels by using the downlink signals in a commercial UMTS network for a pedestrian zone in Nanjing Road, Shanghai.Techniques were developed for extracting the channel impulse responses, and multipath parameters were estimated by using the SAGE algorithm.Channels observed were categorized into LoS and NLoS by using a novel multipath-based LoS detection method.A map overlaid by the channel properties observed at multiple locations demonstrated that LoS channels mostly exist in open areas, while NLoS channels are dominant in the city canyons with densely distributed skyscrapers.Statistical models for the fading coefficients were first established, which have shown that the existence of the LoS path can bring 6 dB gain on average and reduce the gain variation by 5 dB.Additionlly, the statistics of some newly defined parameters were also investigated, which include the distributions of the LoS and NLoS life-distance, the LoS existence probability per user equipment and per NB, power difference at the LoSto-NLoS transition, and the transition duration.Important findings were discovered; for example, the LoS and NLoS life-distances are usually less than 10 m, the LoS-to-NLoS transition can be accomplished less than 10 ms, and so forth.These characteristics are important for the design and preference evaluation of the techniques or systems adaptive to LoS and NLoS scenarios.

Figure 2 :
Figure 2: A photograph of the Nanjing Road.

Figure 3 :
Figure 3: An example of an original channel PDP and the PDP of the channel reconstructed based on the multipath parameter estimates.The powers of the estimated paths versus the path delays are also illustrated.

Figure 4 :
Figure 4: The pdfs of the deviation  between the powers of the strongest path and the first-arrival path in the case where the firstarrival path does not have the largest power among all estimated paths.

Figure 5 :
Figure 5: The locations where the channels are found to be LoS and NLoS channels along the Nanjing Road.

1Figure 6 :
Figure 6: The empirical and fitted cdfs of the channel gains in the LoS and NLoS scenarios.

1
Cumulative distribution functionChannel gain due to shadowing P s (linear)Empirical data for LoSEmpirical data for NLoS - cdf fitted to LoS - cdf fitted to NLoS

Figure 7 :
Figure 7: The empirical and fitted cdfs of shadowing in the LoS and NLoS scenarios.

1Figure 8 :
Figure 8: The empirical and fitted cdfs of multipath fading in the LoS and NLoS scenarios.

7 40Figure 9 :
Figure 9: The empirical pdfs of estimated path loss in LoS and NLoS scenarios.

Figure 10 :
Figure 10: The empirical pdfs of the distance between the NB and the UE in LoS and NLoS scenarios.

Figure 11 :
Figure 11: An example of calculating life-distance of LoS and NLoS channels for a NB with the scrambling code number 188.

3 Figure 13 :Figure 14 :
Figure 13: The empirical percentage of LoS existence probability per location.

Figure 15 :
Figure15:The empirical pdf and fitted Laplacian pdf for the power difference of the channels before and after the LoS-to-NLoS transition.

Figure 16 :
Figure 16: Examples of the four types of transition behavior.

Figure 17 :
Figure 17: The empirical occurrence frequency of transition duration from LoS points to NLoS points or NLoS points to LoS points.
this section, various aspects of the LoS and NLoS channels are analyzed, and the corresponding stochastic models are established based on 664 LoS and 1181 NLoS channel observations.These characteristics include the conventional fading statistics and the distributions of some novel parameters.For notational convenience, we use ℎ LoS, () and ℎ NLoS, () to denote the CIR of the th observation of LoS channel and the th observation of NLoS channel, respectively.The corresponding narrowband channel coefficients ℎ LoS, and ℎ NLoS, are calculated as ℎ LoS, = ∫ℎ LoS, () and ℎ NLoS, = ∫ ℎ NLoS, (), respectively.The channel gains  LoS, and  NLoS, represented in dB are calculated as  LoS, = 20 log 10 |ℎ LoS, | and  NLoS, = 20 log 10 |ℎ NLoS, |, respectively.4.1.Channel Gain, Shadowing, and MultipathFading.The channel gain  LoS, can be written as   =   −   +   +   where   ,   ,   , and   represent the transmission power, path loss, shadowing, and multipath fading denoted in dB, respectively.Figure6depicts the empirical cumulative distribution functions (cdfs) of the channel gains  LoS, and  NLoS, , respectively.It can be observed that the channel gains for LoS are concentrated in the region more to the right of the abscissa than to the NLoS scenario.Different kinds of distributions including the -Nakagami, Rice, Weibull, nor- mal, and - are used to fit with the empirical results.The Kolmogorov-Smirnov test (K-S test) has been conducted for evaluating the consistency of the analytical pdf and the empirical graphs.The results show that both cdfs are well fitted by normal distributions with appropriately chosen parameters reported in Table1.It can be observed from the table that the average of  LoS, is 6.12 dB higher than that in the NLoS

Table 1 :
Parameters of analytical pdfs fitted to the empirical data.
Figure 15demonstrates that the empirical pdf of the power differences Δ LoS−NLoS =  LoS −  NLoS at LoS-to-NLoS transition points.
The fitted analytical graph is also illustrated.It can be observed from Figure15that the majority of the distribution of Δ LoS−NLoS is located above 0 dB in the abscissa.The Laplace distribution with its parameters  = 1.42 dB and  = 2.22 fits well with the empirical data.Furthermore, the observation that Δ LoS−NLoS is less than 0 dB indicates that, for some transitions, the NLoS scenarios have higher received