Research into indoor location-based services has attracted a great deal of attention in recent years. Many of these services have been developed based on trajectory data [1
], such as people’s movement patterns, location prediction, and route planning. Compared to other methods of acquiring indoor trajectories, built-in smartphone sensors provide a more convenient and less expensive method of collecting indoor trajectory data [5
]. However, the trajectory data collected directly from the smartphone are prone to be inaccurate and incomplete, which prevents their immediate use and the analysis of the trajectory data. Thus, more research on indoor trajectory tracking and correction techniques is required.
Map matching has been widely accepted as a common approach for trajectory tracking and correction. By using the road network information as the constraint for vehicles, considerable achievements have been made in the field of outdoor vehicle navigation systems [6
]. Compared with the outdoors, the constraints of the indoor environment are much more complicated. Indoor navigation systems not only need to consider a variety of special elements (e.g., doors, corridors, and stairs) but also need to take more flexible direction information into account. Many indoor matching techniques [9
] have been proposed, including approaches based on particle filter, hidden Markov models (HMMs), and geometric similarity. However, the applicability and accuracy of these methods are still limited. The development of a robust map matching approach for user trajectory tracking that takes into account both accuracy and efficiency remains an open challenge.
Compared with the outdoor environment, indoor space provides richer constraint information. Indoor space consists of a variety of indoor elements, such as rooms, doors, stairs, walls, corridors, floors, etc. The geometric constraints between these indoor elements are far more complicated than outdoor European space and network space. In addition, topological constraints, such as adjacency, association, and inclusion relationships are more common indoors. Based on this, we conclude that the semantic information obtained from the indoor space is more abundant. In general, making use of this information can effectively improve the efficiency of indoor trajectory tracking.
In this paper, we proposed a novel semantic matching method for indoor trajectory tracking. The proposed method is actually an extension of our previous work [5
], which further improves the efficiency and accuracy for user trajectory tracking by combining a semantic-rich link-node model derived from an indoor plan with the semantic information extracted from the trajectory. In contrast to the conventional link-node model, which only considers geometric and topological information, the proposed model further fuses the activity semantics (e.g., “Turn left”, “Turn right”, “Go upstairs”, “Go downstairs”, and “Go into a room”) and direction information (e.g., “East”, “South”, “West”, and “North”) obtained from the floor plan. With the construction of the semantic-rich link-node model, the key to implementing semantic matching lies in the acquisition and recognition of the user’s trajectory data. Pedestrian dead reckoning (PDR), which uses data from built-in sensors in the smartphone to track pedestrians, is a suitable way to acquire trajectory data. In addition, human activity recognition (HAR), which recognizes user activity by sensor data, can be utilized to obtain semantic information about user trajectory.
The remainder of this paper is structured as follows. Section 2
briefly reviews the previous work. Section 3
introduces the link-node model. Section 4
and Section 5
present the semantics extraction steps and the semantic matching process, respectively. Section 6
analyzes and discusses the experimental results. Finally, Section 7
present conclusions and recommendations for future work.
2. Related Work
Map matching is the process of matching trajectory data with the map network, and it can be used to correct the error of the moving object’s trajectory. The information from the map is of great importance to the map matching process, and this aspect has been studied by many researchers. Geometric information such as Euclidean space and geometric shape is the simplest and most commonly used constraint. The main research in this area has taken into consideration road networks [13
] and section shapes [14
]. In particular, the magnetic field information [15
] and direction information [16
] of the indoor building are analyzed and extracted to reduce the PDR error and match the pedestrian trajectory. Moreover, topology information, including the relationship between road networks, was considered in [17
]. In addition, probability information [7
] has also been used when there is uncertainty in the network data. To obtain richer and more comprehensive information, the above information is usually utilized in combination [18
In addition to map information, semantic information has attracted a great deal of attention from researchers. In many cases, semantic information can be used to provide applications that are more meaningful and convenient. To achieve this purpose, Liu et al. [19
] automatically derived semantic locations from the user trajectory and then used them to infer the activities contained in the trajectory. Similarly, Brakatsoulas et al. [20
] constructed a semantic-rich, moving-object model, which was useful for further analysis and data extraction. Spaccapietra et al. [21
] proposed a fully conceptual approach to modeling the moving objects that supported trajectory semantics. In addition, semantic information can also be used for trajectory similarity analysis and location prediction. In [22
], the semantic meaning of the position where the user stayed for a short time was mined to compare the similarity of people’s trajectories. In [23
], the longest common subsequence (LCSS) algorithm was used to determine the semantic similarity. A follow-on study [24
] then considered both geometric similarity and semantic similarity. More semantic information like motion state and mobility context in [25
] can also be of great value to the indoor localization process. Although the semantics of the trajectory have been used in recent studies, the semantic information in the map has not been fully utilized, and how to obtain useful semantic information from the map is worth exploring.
The map matching process can be implemented by a variety of algorithms, ranging from simple search to sophisticated mathematical tools. The main purpose of these algorithms is to improve the accuracy and efficiency of the matching process. According to the global optimal idea (i.e., the cumulative error is minimal), methods based on the HMM [26
] or interactive voting [27
] can be used to improve the accuracy of the map matching. On the other hand, the matching efficiency can also be optimized by incremental calculations [8
], but this will reduce the accuracy. In addition, some complex machine learning methods and data mining methods (e.g., the Kalman filter [28
], fuzzy theory [29
], and artificial neural networks [30
]) can also be utilized to address the problem of map matching. Although these algorithms can achieve a high precision in a particular situation, they require a large quantity of data to be trained. In general, an efficient and highly accurate matching method still needs to be further studied.
In an indoor setting, the map matching process involves matching the pedestrian trajectory to the floor plan. Although the indoor environment is more restrictive, it also brings more useful information. For example, walls and corridors can provide geometric information to limit the user’s movement [9
], and doors, stairs, and other indoor elements are key information for map matching [12
]. Walder et al. [31
] described in detail how these elements were used to correct position. These features can also be used as landmarks, for example, Wang et al. [32
] used mobile devices to sense indoor landmarks, which can be used to recalibrate pedestrian’s indoor location. Compared to the outdoor environment, indoor map matching algorithms are more complex. Particle filter-based approaches have been commonly used to match indoor maps. In [33
], the user’s step length and heading direction were integrated with the floor map by a particle filter algorithm. In further research [34
], Wi-Fi, map information, and the inertial data were combined by a particle filter to achieve pedestrian tracking in an indoor environment. However, approaches based on particle filters require significant computational resources. To reduce computational complexity, Xiao et al. [36
] used a conditional random field model, which combines maps and state observations, to track indoor pedestrian.
In our previous work [5
], we combined PDR, HAR and landmarks to obtain semantic-rich indoor trajectory data. However, the previous method uses geometric distance to match the landmark and correct the trajectory, which is still undesirable in both computation efficiency and error elimination for applications. In order to overcome these problems, a new semantic matching method is proposed, and some optimization measures are used in the matching process. In detail, a link-node model is first abstracted from the floor plan, and rich navigation-related semantics are incorporated into the model. The semantics extracted from the trajectory are then used to match the nodes in the link-node model. The matching method can not only efficiently track pedestrian trajectories, but it can also accurately describe them.
3. Semantic-Rich Indoor Link-Node Model
The link-node model [37
] has a significant advantage in the expression of indoor spaces. Since special attribute points and points of interest (POIs) can be easily abstracted from the map, and people usually walk along the center line between two nodes in a narrow space, a link-node model is constructed, as shown in Figure 1
. The nodes are normally door points, stair points, and turning points, and links are composed of starting nodes and ending nodes. To simplify the proposed model and increase its computational efficiency, some nodes with the same semantics are abstracted as a flag.
indicates the adjacent doors (
) on the south side of the corridor,
represents doors (
) on the north side of the corridor,
represent the east doors (
) and the west doors (
), respectively. Similarly,
indicate the corresponding nodes on the second floor.
represent the turning nodes in the corridors.
Semantic information can not only express the attributes of a node but can also describe the pedestrian movement behavior of the corresponding links. These semantics are of great value for the description of the indoor location and the pedestrian trajectory tracking. It therefore makes a lot of sense to integrate semantics into the model and to use this for pedestrian trajectory matching. In order to correspond to the pedestrian’s indoor navigation elements, attributes such as “Turn”, “Stair”, and “Door” are assigned to each node, and semantics such as “Turn right”, “Turn left”, and “Go straight” are added to each link. When the straight line distance is long enough (>5 m), the “Go straight” link is added; otherwise, the semantics are left empty (). In addition to the basic semantic information, distance and direction information (e.g., “North”, “East”, “South”, and “West”) is also added to further enrich the proposed model.
describes the semantic-rich link-node model, in which a semantic node is determined by five basic information elements. The Id is the identifier of a node. An attribute is a set of stairs, a turn, or a door. Adjacent links present the distance, direction, and semantics between the current node and the next nodes. Direction information indicates the change of direction when the user passes through the node. Similarly, the semantic description represents the detected semantic information in this process. Figure 3
shows the semantic-rich information of node
as an example.
In this paper, a semantics-based map matching method was proposed for accurate and efficient indoor user trajectory tracking. Differing from the previous methods, we provided an appropriate formalism to model the rich semantics related to indoor navigation systems. Thus, a semantic-rich indoor link-node model was constructed from the floor plan and used as the map basis for the following trajectory matching. For the acquirement of the continuous indoor pedestrian trajectory, smartphone-based PDR technology was applied. Given the collected trajectory data, an HAR approach was adopted to identify the user indoor activity. Based on this information, a trajectory could be logically segmented and the semantic information and distance information could be inferred. The inferred information was finally utilized to match the inaccurate trajectories with the proposed link-node model. In addition, the matching process was accelerated by first using the most important semantic information. The conducted experiments confirmed that the proposed method can achieve an accurate trajectory tracking result while guaranteeing a high matching efficiency.
In practical applications, buildings serve different functional purposes and may have different layouts and structures. The motion patterns among different users may also show large differences. In order to further improve the applicability of the proposed method, more complex indoor environments and more types of trajectories will be considered in our future work. Furthermore, extending the proposed semantics-based trajectory tracking to real-time applications is a potential direction for further improvement. It is also an interesting research area to restore the missing indoor features by analyzing the trajectory data.