The iParking service aims at creating a smoother and more efficient packing process, enhancing the utilization efficiency of parking spaces and protecting the environment. Related stakeholders include parking users, i.e., drivers, owners and operators of parking facilities, and the environment, such as air quality and traffic conditions that have direct effects on the public. Customers need be able to know the availability of parking spaces in advance and reserve a parking place before their trips. They also require intelligent services for activities after parking, e.g., navigation toward places of interest, providing intelligent information and finding routes backward their own cars after they have completed their activities. Owners and operators of parking facilities expect to improve the parking space utilization efficiency, and as well to provide a pleasant customer-user experience, and promote their business. Urban authorities, who stand for the public, require less unwanted driving to reduce traffic congestion and environmental pollution. As a summary, a new generation intelligent parking service is required to provide all of the stakeholders with the following advantages:

Efficient and Proactive Mobility Management. An effective parking solution plays a critical role in encouraging the utilization of more efficient transportation means, and it is effective way to alleviate problems such as traffic congestion, roadway costs, pollution emissions, energy consumption and traffic accidents. It is important for users to be able to manage their daily mobility in a proactive way.

Based on these user requirements, an intelligent parking service should provide customers with functionalities for all three stages of a transportation process, including transport planning, driving, and activities after parking. A parking facility monitors the occupancy status of its parking spaces using vehicle detection sensors, and updates online the occupancy status information through a database. In addition, the database system manages the reservation requests of customers and displays on-site the reservation status of each parking space. Mobile Internet makes the service ubiquitously accessible. The iParking service currently provides four types of functionalities as follows:

The iParking service is implemented through a distributed architecture as shown in Figure 1. The Internet connects the different parties, including parking facilities, customers and service providers. The service providers establish the interfaces to bridge parking facilities and users over the Internet, and develop tools to create and maintain the parking facilities database.

Operators of parking facilities use sensors to monitor the occupancy status of each parking space and publish online the parking space occupancy status information in real time. At the same time, the database system handles users' request for parking space reservations, and displays the on-site reservation status to prevent occupancy by others. Every parking facility manages its own database, which consists of information related to the whole parking facility and each of parking spaces. Information related to a whole parking facility includes, for example, permitted parking periods and other general messages, while the information of each parking space includes its geographical coordinates, occupancy status, reservation status, etc. At the user end, a client program runs on his smartphone to access the service through the Internet. The client is a light-weight application as a user only collects the databases of the parking facilities of his interest.

User operations of the iParking service are conducted on the iParking client software, which has a graphical interface as shown in Figure 2. In its current form, the program is developed with a NOKIA N8 smartphone, which runs with Symbianˆ3 operating system (OS). The software development is done with Qt SDK and the Qt Creator integrated development environment (IDE) [27]. QtMobility application programming interfaces (APIs) are used to acquire the measurements of the smartphone sensors for indoor positioning. NOKIA maps are employed for on-road navigation, while an image-based floor map is used as indoor map.

WLAN RSSI signals have a statistical dependency on geographical locations. The location-RSSI dependency has been widely used for indoor positioning as this kind of positioning solution is cost-efficient and they can be operated in conjunction with communication services. Unlike traditional solutions, which usually use a specific hardware tag for positioning, the proposed HIPE utilizes the built-in hardware of a smartphone to collect the observables and performs position estimation using the smartphone's computational resources. Significant advantages of re-using a smartphone platform include that the positioning solution is more cost-efficient, and it is more convenient to integrate the positioning solution with related applications. However, a disadvantage is that the development of the HIPE is restricted by the API availability of the smartphone platform. For example, the API of Symbianˆ3 operating system used in this paper restricts the interval of two WLAN scans to 8–10 s for energy saving purposes [31]. That means a user will not receive an updated RSSI positioning estimate during such intervals, and this causes an unsatisfactory experience. Therefore, the HIPE propagates locations using the DR technique during the interval of two WLAN scans.

The location-RSSI dependency is learnt and recorded as prior knowledge in a database. Due to the highly non-stationary nature of WLAN signals, RSSI observables at a given location usually have a variance following a non-Gaussian, left-skewed distribution [5]. In this study, a parameteric Weibull function is used to represent the probability density function (PDF) of RSSI observables at a given location, and its three parameters are calculated using a set of RSSI observables [5,31]. It was concluded in [5] that the utilization of the Weibull probability density function can reduce the work load by using a lesser number of RSSI observables than the traditional histogram-based method, and it can improve the positioning accuracy by 20% compared to the histogram-based method when a same number of RSSI observables were used. Given a predetermined grid of reference points (RP), RSSI measurements of multiple WLAN access points (AP) are collected at each RP to calculate the probability distribution of RSSI with the Weibull function. As a result, the learning course produces a database of the probability distributions: (1) ∑ = { P ( O | X = S i ) }where P(O|X = S_i) is the probability in which a vector of RSSI observables O = {o^k}(k = 1, ⋯, M) is collected at a given RP S_i (i = 1, ⋯, N), M is the number of APs and N is the number of RPs.

With the HMM methodology, the HIPE solution considers the motion of a user as a temporally correlated process, instead of a series of isolated points, and it estimates positions using the solutions of HMM problems such as the grid-based filter. The HMM methodology has been presented with details in [6,32], and an implementation of the proposed HIPE has been described in [5,6,33], which included the methods of determining the HMM parameters (A,B,π).

The presence of walls, humans and other rigid objects indoors causes multipath and non-line-of-sight propagation of wireless signals, and results in high variance of WLAN RSSI observables that degrades positioning accuracy [34]. Furthermore, a WLAN scanning probably fails to obtain RSSI observables from one or multiple APs due to environmental disturbances or excessive distances between the smartphone and APs. Therefore, it is important to monitor the integrity of RSSIs to improve the positioning accuracy and reliability. In the proposed HIPE solution, two physical natures of RSSI scanning are relied upon in the RSSI integrity monitoring [5]:

Proposition 1: When a scanning mobile device is far away from an access point and its RSSI observable is lower than a threshold, the RSSI variance and the rate of scanning failure are significantly higher.

Proposition 2: The probability distribution pattern of valid RSSI measurements is most matched with that of a given location in the knowledge database when and only when the given location is a true position where RSSIs are observed.

In the proposed HIPE solution, the threshold in Proposition 1 is defined as −75 dBm, and valid RSSI measurements are defined as those observables which are acquired successfully by a scan and are stronger than the threshold. The degree of pattern matching between observed RSSIs and the prior knowledge is numbered by an emission probability, which is a joint probability of all involved measurements and is calculated as follows: (2) b i ( t ) = P ( o t 1 | X ( t ) = S i ) ⋯ ∗ P ( o t k | X ( t ) = S i ) ⋯ ∗ P ( o t M | X ( t ) = S i ) where S_i stands for i-th reference point (1≤i≤N), t is epoch number, o t k is the observation of RSSI of k-th access point at epoch t, P ( o t k | X ( t ) = S i ) is the probability in which RSSI of o t k is observed at a given location S_i, it is obtained from the knowledge database of the location-RSSI dependency.

In the integrity monitoring process, an emission probability is first calculated for each of RPs using RSSIs of all APs, including those of scanning failures, which are marked as “0”. The maximum emission probability and the corresponding RP are recorded as a benchmark. In this case, a scanning failure is considered as useful information that the scanning device is located far away from the corresponding AP with a high probability. Second, valid RSSI measurements are selected to calculate an emission probability for each of RPs. The maximum emission probability of valid RSSI measurements is compared with the benchmark of all RSSIs, and the significance of the difference is evaluated using multiple elements such as the numbers of successful and failing scanning APs respectively, and the numbers of RSSI measurements which have high and low probabilities P ( o t k | X ( t ) = S i ), respectively. If the two sets of emission probabilities are coherent, the emission probabilities of valid RSSI measurements are used to calculate a position estimate. Otherwise, the emission probabilities of all RSSIs are used, and as a result the derived position estimate may suffer major errors with a high probability.

Relative positions measured by the smartphone sensors are used to propagate positions with the DR technique during a long RSSI scanning interval. In the parking application, relative locations need to be estimated in two motion modes: vehicle mode and pedestrian mode. Vehicle mode is a platform-fixed mode, and it is used for in-vehicle navigation before parking, while pedestrian mode is used for pedestrian navigation and guidance in activities after parking. Different algorithms are used in the different modes. In vehicle mode, the smartphone compass and accelerometers comprise a classic inertial navigation system (INS). Given a fixed body frame, a compass measures motion heading, and measurements of built-in accelerometers are integrated once to produce the velocity, and twice to produce the travelled distance [6]. Consequently, relative location is calculated in a local coordinate system. On the contrary, the integration operation is not applicable in pedestrian mode as it is difficult to keep a smartphone in a fixed body frame. Instead, a pedestrian distance is estimated through accumulating step lengths of occurred steps, which is widely used in pedestrian dead reckoning (PDR) techniques [4,35–38]. Based on the pedestrian acceleration pattern, the occurrences of steps are recognized using accelerometer measurements, and the lengths of all recognized steps are accumulated to calculate moved distance. In this paper, we use a constant value of 0.7 m per step to estimate step lengths. With the estimate of travelled distance (d_t) and heading (α_t) during an interval, relative motion vector D_t is represented in a horizontal plane as follows: (3) D t = [ d t cos ( α t ) d t sin ( α t ) ]where t is an epoch number.

Thereafter, propagated positions in the horizontal plane are calculated based upon a previously determined position as follows [39]: (4) X n = X 0 + ∑ t = 1 n D twhere t is an epoch number, X₀ is a previously determined position, D_t is the relative position during a propagation interval. In this study, the propagation interval is 1 s, and position estimate is updated every second through the DR technique during a WLAN scanning interval of 8–10 s. Thus, the availability of positioning solution is enhanced.

This section presents the experiment results concerning positioning accuracy and availability. In the experiments, we used NovAtel GPS/IMU integrated SPAN system to obtain a trajectory reference. However, as specified in its manual, the SPAN system has increasing errors over time when GPS does not work well indoors. In order to provide the SPAN system periodic references, we recorded nearest true reference points every 30 s during the experiment, and calibrated positioning results of the SPAN system in post mission using the recorded positions of RPs. The calibrated SPAN positioning results achieved decimeter-level accuracy, and were used as the reference of error calculation. Positioning error was calculated epoch by epoch as follows: (5) ɛ i = ‖ P i − P ¯ i ‖where P_i is the HIPE positioning result; P̄_i is the reference position.

The experiment lasted for more than 2,000 s, and it obtained 208 RSSI positioning results as the smartphone platform restricts a WLAN scanning interval of 8–10 s. For performance comparison, the recorded RSSI measurements were processed also with the classic Maximum likelihood estimation (MLE) algorithm. The same integrity monitoring processing was applied for the two algorithms, i.e., the grid-based filter and the MLE algorithm. In the HMM approach, the location of the first epoch was estimated using the MLE method, and it was used to calculate the initial probability distribution π [6]. The transitional probability matrix A was calculated using the measured movement distance and heading, and accordingly the ratio K between high and low transitional probabilities adaptively varied at K ∈ [2000, 200000] [30]. Given an update interval of 1 s, the dead reckoning approach propagates positions during two scanning epochs based on a previous position estimate of the grid-based filter.

Table 1 compares the positioning errors of the three methods in terms of RMS error (RMSE), average error (AE) and 95% radius error (R95). The grid-based filter has a significantly reduced RMSE by 1.62 m (32.7%), AE by 1.65 m (47.7%) and R95 by 6 m (50%) than the MLE.

Figure 4 shows the time series of positioning errors of the three solutions. It further indicated that the grid-based filter solution has much less errors than the MLE solution, and the DR-integrated solution has an update rate of 1 s, which indicates a higher availability compared to the update rate of 8–10 s in the grid-based filter and MLE solution.

Errors of the DR-integrated solution are dominated by those of the grid-based filter solution as the dead reckoning is based on a latest result of the grid-based filter. However, large errors of the grid-based filter solution are mitigated in the DR-integrated solution as the quality of a new RSSI positioning estimate has been checked using the DR result and large outliers are mitigated in the final DR-integrated solution. The results as shown in Table 1 and Figure 4 indicate the DR-integrated solution, which is the final HIPE solution, has positioning errors of less than the width of a parking space (2.9 m) in most epochs, and it can effectively support the location-based functionalities in the iParking service.