FloorTag: A Hybrid Indoor Localization System Based on Floor-Deployed Visual Markers and Pedometer Integration

Abstract

With the widespread adoption of smartphones and wearable devices, localization systems have become increasingly important in modern society. While Global Positioning System (GPS) technology is widely accepted as a standard outdoors, accurately determining user location indoors remains a significant challenge despite extensive research efforts. Indoor positioning systems (IPSs) play a critical role in various sectors, including retail, tourism, transportation, healthcare, and emergency services. However, existing solutions require costly infrastructure deployments, complex area mapping, or offer suboptimal user experiences without achieving satisfactory accuracy. This paper introduces FloorTag, a scalable, low-cost, and minimally invasive hybrid IPS designed specifically for smartphone platforms. FloorTag leverages a combination of 2D visual markers placed on floor surfaces at key locations, and inertial sensor data from mobile devices. Each marker is associated with a unique identifier and precise spatial coordinates, enabling an immediate reset of accumulated localization errors upon detection. Between markers, a pedometer-based dead reckoning module maintains continuous location tracking. The localization process is designed to be seamless and unobtrusive to the user. When activated by the app during navigation, the phone’s rear camera, naturally angled toward the floor during walking, captures markers. This solution avoids explicit user scans while preserving the performance benefits of visual positioning. To model the indoor environment, FloorTag introduces the concept of Path-Points, which discretize the walkable space, and Informative Layers, which add semantic context to the navigation experience. This paper details the proposed methodology and the client–server system architecture and presents experimental results obtained from a prototype deployed in an academic building at the University of Catania, Italy. The findings demonstrate reliable localization at approximately 2 m spatial granularity and near-real-time performance across varying lighting conditions, confirming the feasibility of the approach and the effectiveness of the system.

1. Introduction

The last decade has seen a shift in how individuals interact with information and their environment, largely driven by the ubiquity of smartphones and pervasive internet connectivity [1,2]. These devices, with their ever-increasing array of sensors and computational power, act as gateways between the physical and digital world, enabling a new set of emerging applications and services which tailor information and functionalities based on the user’s position (location-based services, LBS). While outdoor LBS, primarily due to the availability and reliability of the Global Positioning System (GPS), have experienced exponential growth, GPS signals are often attenuated or completely blocked by building structures (roofs, walls), making the technology ineffective indoors and limiting the adoption of LBS in indoor environments [3]. The absence of a de facto standard for indoor positioning as high-performing as GPS has raised significant research and development efforts. Various methodologies and technologies have been proposed, each with its own set of advantages and disadvantages in terms of accuracy, cost, infrastructure dependency, scalability, and user experience [4]. According to Business Research Insights (Business Research Insights, available at: https://www.businessresearchinsights.com/market-reports/indoor-positioning-and-navigation-systems-market-106632, accessed on 7 October 2025), the indoor positioning systems (IPSs) market size is projected to reach USD 21.31 billion by 2033 with a compound annual growth rate (CAGR) of 24.5%, underscoring the substantial commercial and societal value of effective indoor navigation solutions. These systems are increasingly adopted across diverse domains such as smart buildings, healthcare facilities, retail environments, industrial warehouses, museums, and logistics hubs, where location-aware services enhance operational efficiency, safety, and user experience. In the context of last-mile delivery (LMD) [5], efficient indoor localization represents an important factor for reducing service time (the time couriers spend locating the correct recipient or delivery point within large, multifloor indoor environments where signage is often insufficient). Integrating lightweight indoor positioning systems into LMD workflows can accelerate package handovers and improve overall operational efficiency [6]. Despite the advancements in the field, no single solution has emerged as widely effective. Many systems rely on dedicated hardware infrastructure, such as Wi-Fi access points or Bluetooth beacons, resulting in high deployment and maintenance costs [7]. Other approaches, such as those based on magnetic fingerprinting [8] or visible light communication (VLC) [9], require extensive pre-mapping or environmental modifications. Vision-based methodologies, while promising due to the high-quality cameras in modern smartphones, often demand active user involvement (e.g., scanning markers on walls), leading to poor user experience [4]. An ideal IPS should satisfy several key requirements [10,11], such as the following:

• Low cost: The system should avoid expensive or complex hardware infrastructures, thereby reducing barriers to adoption.
• Low invasiveness: Environmental modifications should be minimal, preserving visual integrity of spaces and ensuring compatibility with existing systems.
• Simple deployment: Installation should be feasible by non-specialized personnel without requiring extensive technical expertise, directly impacting scalability and adoption potential.
• User-friendly: The system must provide intuitive interfaces with passive localization and clear navigation guidance, minimizing cognitive workload for end users.
• Accuracy: Positioning performance should remain reliable across varying environmental conditions, meeting application-specific requirements.
• Scalability: The architecture must integrate seamlessly into commercial solutions and adapt to different building types and sizes with minimal customization effort.

However, existing solutions do not meet these criteria, often requiring costly infrastructure deployments, complex area mapping, or active user involvement, while failing to consistently achieve satisfactory accuracy. This paper addresses these limitations by proposing FloorTag, a novel hybrid IPS that combines computer vision for absolute positioning via strategically deployed 2D visual markers with inertial sensor-based dead reckoning to maintain tracking between markers. Each marker is associated with a specific position on the building’s map. The key contributions of this work include the following:

• A novel indoor localization methodology based on floor-deployed markers (combined with pedometer-based dead reckoning) which, unlike solutions requiring wall or ceiling scans, leverages the natural downward orientation of a user’s smartphone camera during navigation. This makes the marker detection process nearly transparent and unobtrusive to the user while walking and minimizes infrastructure needs, requiring only printed markers. The system achieves near-real-time performance (see Section 4 for details) on older smartphones and does not require internet connectivity for its core localization functions.
• The introduction of Path-Points, a concept for discretizing the walkable space into predefined locations. This serves to stabilize position estimates and mitigate sensor drift.
• The introduction of Informative Layers, which overlay semantic information (such as room labels and points of interest) onto the base coordinate map of the indoor environment, enhancing the user experience by enabling goal-oriented navigation beyond coordinate-based navigation.
• A Client–server architecture that ensures system scalability by centralizing building data management on the server while delegating real-time processing tasks to the client devices.

This specific focus on optimizing the trade-off between minimal deployment effort and cost-effective scalability, passive user involvement, and navigation expressiveness remains largely unexplored in previous hybrid, marker-based indoor localization research.

The remainder of this paper is organized as follows: Section 2 reviews related work on IPS technologies, highlighting strengths and limitations of existing approaches. Section 3 describes the FloorTag system architecture and its key components. Section 4 presents experimental results from a prototype deployment in an academic building at the University of Catania, Italy. Finally, Section 5 summarizes the contributions and outlines future research directions.

2. Related Work

Indoor positioning has been studied for three decades, resulting in a wide range of approaches, datasets [12], and systems, each with specific strengths and limitations, aimed at providing accurate and reliable indoor localization. Key technologies include radio frequency (RF), magnetic field, acoustic, inertial, and vision-based methods [4,13,14].

RF-based approaches are among the most commonly adopted in IPSs. They leverage the properties of radio signals to estimate a user’s position within complex indoor environments. Common RF communication technologies include Wi-Fi, Bluetooth Low Energy (BLE), Radio Frequency Identification (RFID), and Ultra-Wideband (UWB) [7,15,16].

Wi-Fi-based IPSs are extensively used due to the pervasive availability of Wi-Fi infrastructure in most buildings [17]. They typically rely on Received Signal Strength Indicator (RSSI) fingerprinting [18], trilateration [19], or triangulation [20,21], achieving wide coverage and relatively low deployment costs but often limited accuracy. BLE IPSs provide accuracy levels comparable to Wi-Fi-based systems, supporting trilateration [22], proximity sensing [23], and RSSI fingerprinting [24]. Their low power consumption makes them suitable for smartphones and wearable devices. Despite these advantages, BLE-based systems require dedicated infrastructure and, like Wi-Fi solutions, may not meet the accuracy requirements of all indoor LBSs. RFID technologies employ tags (active, passive, or semi-passive) that communicate with readers via radio waves, enabling real-time localization with high accuracy in short-range scenarios [25]. While RFID solutions benefit from their small tag size and low energy requirements, advanced algorithms, such as Kalman filters [26] or K-nearest neighbor (KNN) [27], are often necessary for accurate position estimation. UWB systems utilize a wide radio spectrum to transmit low-power, high-bandwidth signals over short distances, exhibiting high resistance to multipath effects and obstacle penetration capabilities, which makes them particularly suitable for industrial applications [28]. They achieve very high accuracy but their implementation is more complex due to the need for specialized hardware and precise signal acquisition and synchronization [29]. In terms of positioning accuracy, Wi-Fi-based systems typically range from 5 to 15 m, BLE systems from 2 to 5 m, and RFID systems from 1 to 5 m. UWB represents the most accurate RF-based technology, often reaching decimeter- or even centimeter-level precision [30].

Magnetic-field-based technologies have gained significant attention for indoor localization due to the ubiquitous presence of magnetic fields. Geomagnetic anomalies within buildings, caused by structural elements and electronic equipment, create distinctive patterns that can be exploited for positioning purposes [8]. Unlike Wi-Fi signals, magnetic field measurements exhibit high temporal stability, uniqueness due to ferromagnetic disturbances (e.g., from steel reinforcements), and robustness to environmental changes such as human movement. Moreover, magnetic-field-based approaches represent a low-cost solution, as they leverage sensors already embedded in commercial smartphones and require no additional infrastructure. However, the creation of magnetic fingerprint maps is labor-intensive, often requiring extensive on-site data collection. Additionally, magnetic field measurements may suffer from low spatial discernibility (similar readings occur at different locations) and from sensor heterogeneity across smartphone models, which can lead to inconsistent results [31]. Several studies have investigated magnetic-field-based localization, proposing different techniques for fingerprint acquisition, matching algorithms, and map construction strategies to address these challenges and improve system performance [32,33,34,35,36].

Sound-based technologies, including both audible and ultrasonic waves, have also been explored. One of their advantages lies in the relatively low speed of sound compared to electromagnetic waves, which simplifies time synchronization and improves the accuracy of time-based ranging methods.

Ultrasonic approaches typically operate at frequencies above 20 kHz and estimate distances using various techniques, primarily time-of-flight (ToF) and time difference of arrival (TDoA), between transmitters and receivers [37,38]. These systems can achieve centimeter-level accuracy under controlled conditions. However, they are susceptible to multipath propagation, signal attenuation due to obstructions, and interference from environmental noise. Audible-sound-based systems utilize ambient acoustic signals captured by built-in smartphone microphones. These can include controlled sounds (e.g., emitted by fixed loudspeakers in malls, consumer stores, or museums) [39,40] or uncontrolled sources (e.g., environmental audio used for fingerprinting) [41]. Such methods are cost-effective, as they rely on existing hardware and infrastructure. Nevertheless, their accuracy may be affected by dynamic environmental conditions, user orientation, and background noise. Hybrid systems combining audio with other sensor modalities have been proposed to mitigate these limitations and enhance robustness [42,43].

Dead-reckoning-based IPSs leverage a smartphone’s inertial sensors (accelerometer, magnetometer, and gyroscope) to estimate the user’s position starting from a known point. These systems offer advantages including low cost, small size, and low power consumption. Additionally, they do not require external infrastructure or installation, as they rely solely on onboard sensors. While it is theoretically possible to achieve precise tracking by computing velocity and direction, in practice, sensor drift quickly introduces significant cumulative errors, making these approaches unsuitable for long-term localization without external corrections [44]. To address this limitation, step-counting algorithms and orientation estimation techniques are typically employed [45,46], often combined with other sensor-based approaches that provide periodic recalibration to reset the accumulated errors [47,48]. Comprehensive surveys on inertial navigation approaches can be found in [49,50].

Hybrid approaches combine data from diverse sources, such as inertial, RF, magnetic, visual, and environmental sensors. Even though single-sensor approaches have demonstrated promising outcomes in specific scenarios, the variability and dynamism of indoor spaces often require the integration of multiple sensing modalities to exploit their complementary strengths and achieve robust and accurate localization [51]. By fusing heterogeneous data, these systems can overcome the limitations of individual technologies, leading to significant enhancements in both accuracy and reliability, particularly within complex and variable indoor settings [52,53]. For a comprehensive overview of hybrid localization techniques, the reader is referred to dedicated surveys [54,55].

Camera-based methodologies leverage the high-resolution cameras embedded in modern smartphones to implement various vision-based indoor localization appro- aches [56,57]. These can be broadly classified into VLC, computer vision, and visual-marker-based techniques.

VLC systems utilize smartphone cameras to detect and decode high-frequency light signals modulated by LEDs. Such signals, imperceptible to the human eye, carry positional information for localization [58,59,60]. By exploiting the rolling shutter effect of CMOS cameras, these systems can decode signals in a single frame and achieve decimeter-level accuracy due to light’s directional properties [61,62]. VLC approaches offer advantages such as operating in an unlicensed spectrum band and enabling spatial reuse of light; however, they require direct line of sight and their performance can be affected by ambient lighting conditions. Moreover, they can be expensive to deploy if suitable LED infrastructure is not already in place and are energy-intensive due to continuous camera usage [9]. Computer-vision-based localization techniques exploit the smartphone’s camera to acquire detailed visual information from indoor environments. They match live camera views to a pre-built or online visual map using feature matching [63,64], photometric alignment [65], or learned models to estimate the user’s pose (position and orientation) [66,67]. To improve robustness and achieve fine-grained accuracy, these techniques often fuse visual data with on-device inertial measurements [68].

Visual-marker-based approaches rely on a set of predefined markers, which are easily detectable and strategically distributed across the indoor environment. Computer vision algorithms detect and decode the unique identifiers encoded within these markers, which are then matched against a database of known marker positions to determine the user’s precise location in the mapped indoor space. Furthermore, the relative pose of the camera can be usually computed from the marker’s image, obtaining both position and orientation of the device. Marker-based systems are commonly integrated into hybrid localization approaches, where they serve as reliable reference points to reset drift and cumulative errors from other positioning methods, thereby maintaining long-term accuracy and preventing the degradation of positioning precision. Various implementations of marker-based approaches have been proposed, employing different types of visual markers and integration strategies. In [69], Idrees et al. presented an Android indoor navigation prototype for blind users that relies on QR code strips placed across corridor floors to provide discrete localization at each scan; the app offers text-to-speech and gesture-only interaction, computes routes (shortest or fewer-turns) via the A* pathfinding algorithm, and validates progress at each scan to detect and correct deviations. However, covering corridors with continuous strips results in high visual invasiveness, and, technically, standard QR codes are less robust to motion blur and slower to decode than other types of markers, preventing fluid detection while moving. Moreover, the system provides only discrete localization updates, lacking the continuous tracking and orientation stability required for seamless navigation between markers. Khan et al., in [70], proposed a low-cost, generic smartphone indoor navigation system that uses simple ARToolKit fiducial markers printed on paper and mounted on ceilings; the phone’s camera detects each marker to determine the user’s position, infers user direction from camera–marker axis alignment, and provides audio/text guidance. Ceiling placement reduces occlusions and keeps markers within the camera’s field of view, enabling robust detection under typical indoor lighting. Nevertheless, this configuration faces practical limitations. Scaling the approach to high-ceiling venues requires increasing marker dimensions, resulting in visual intrusiveness, or is sometimes unfeasible—such as in large commercial centers—where the excessive ceiling height prevents reliable marker decoding or requires markers that are too large. Installation on the ceiling also adds significant deployment effort. Furthermore, the absence of integration with other sensor fusion-based approaches requires a high marker density whenever satisfactory accuracy is needed, while the requirement to point the device upwards may prove ergonomically unnatural during navigation. Ruvolo et al., in [71], described a smartphone-based indoor mapping and navigation system that combines visual–inertial odometry with sparse Apriltag markers to build building-scale maps and localize the user. The method optimizes tag and trajectory poses, then performs navigation over a path graph derived from user walks. This hybrid reliance on visual–inertial odometry and graph-based optimization is likely to be more demanding in terms of computation and battery than minimal, purely marker-based pipelines. Additionally, the mapping process is labor-intensive, requiring a complete physical traversal to construct the path graph, unlike systems that allow discrete, plug-and-play deployment. Neto et al. [72] introduced a tape-shaped fiducial marker designed for indoor navigation and localization systems, leveraging computer vision to enable continuous reading along environmental perimeters. The marker features hierarchical coding patterns for multiscale detection, minimizing service interruptions, and includes a web application for generating markers from floor plans. Evaluations in 3D simulations and real-world smartphone tests demonstrate superior performance over ArUco, QRCode, and STag in multidistance detection, with robustness to lighting variations, viewing angles, and partial occlusions, though sensitive to blurring and with higher detection latency, which can compromise the fluidity needed for real-time, handheld navigation. Furthermore, the requirement to deploy a continuous tape along the perimeter of the environment imposes significant installation effort and introduces visual intrusiveness, particularly in existing buildings.

The literature review shows that a wide range of indoor navigation systems have been proposed, each targeting specific requirements but also exhibiting different limitations. While recent solutions can deliver accurate and robust localization under controlled conditions, practical deployment still faces important challenges in terms of infrastructure and installation cost, user effort and usability, configuration and maintenance complexity, and scalability to heterogeneous building layouts. In this context, FloorTag, by shifting the paradigm from active scanning to passive visual interaction, specifically addresses the trade-off between deployment simplicity and operational robustness, ensuring a scalable and privacy-preserving architecture that minimizes the cognitive burden on the user. Furthermore, the discretization of walkable space into Path-Points optimizes pathfinding computational efficiency and ensures trajectory consistency, while the integration of Informative Layers elevates the system to a context-aware navigation assistant capable of semantic, goal-oriented guidance.

3. System Overview and Architecture

The FloorTag IPS adopts a hybrid approach that combines visual marker detection with pedometer-based dead reckoning. The key design choice is to strategically place visual markers on the floor, rather than on walls or other vertical surfaces, motivated by the observation that smartphone users typically hold the device with the rear camera oriented downward during navigation. Moreover, the floor offers intrinsic features that improve detection performance:

• The floor typically presents a consistent and predictable background pattern, as shown in Figure 1. This characteristic can be exploited to improve decoding algorithm speed and allows reduction of physical marker sizes.
• The marker appears with a relatively fixed and known size within the camera frame due to the stable distance between the smartphone camera and the floor during typical usage. This characteristic enables faster marker detection within the frame.
• The predetermined positions and spatial relationships of markers in the building layout enable sequence validation: given a detected marker, subsequent detections are constrained to spatially adjacent positions, allowing the system to reject spurious detections caused by decoding errors, improving overall reliability.

The system employs AprilTag markers [73,74] for accurate indoor positioning, as they are reliably and quickly detected even by older smartphones. It uses the AprilTag 2 detector, which improves robustness and efficiency over the original implementation while retaining the same tag families and coding scheme. AprilTags demonstrate robustness to illumination changes caused by dynamic environments (characterized by the presence of other people, on–off switching of lights, and shadows) and maintain detection capabilities for blurred or out-of-focus markers resulting from rapid user movements. Detection remains effective even with small marker sizes at typical handheld distances during walking, across a variety of floor surfaces (light, medium, and dark) and lighting conditions [75] (details in Section 4). The availability of the source code allows developers to modify the detection algorithms to better adapt to specific floor characteristics.

To maintain position estimation between marker detections, the system integrates pedometer-based dead reckoning. This hybrid design addresses the limitations of both individual approaches: visual markers provide accurate absolute positioning but require a line of sight, while pedometer-based tracking ensures continuous localization but suffers from cumulative drift over distance. Figure 1 illustrates the concept.

FloorTag relies exclusively on standard smartphones, including older models, and passive visual markers, eliminating the need for costly infrastructure deployment. Marker dimensions and placement strategies were optimized to minimize visual impact while maintaining detection performance. Furthermore, the system architecture was designed for ease of deployment, allowing non-technical personnel to place the markers and manage system setup effectively. To ensure an intuitive user experience, marker detection operates passively without requiring explicit scanning actions, addressing a key limitation of conventional marker-based approaches that demand active user engagement to locate and scan markers. Finally, the system is scalable by design, as it is straightforward to extend marker coverage to larger environments or integrate additional indoor positioning technologies. Before detailing the FloorTag’s architecture, the concepts of Path-Points and Informative Layers must be introduced.

3.1. Path-Points

In structured indoor environments, pedestrian movement is typically constrained to narrow pathways such as corridors and aisles, and continuous position tracking translates short-term sensor noise and step-length variability into visually irregular trajectories without adding useful navigational information. This degrades the user experience and may produce geometrically impossible locations, such as positions appearing through walls, due to accumulated pedometer drift. To address this issue, the walkable space is discretized into a set of Path-Points $\mathcal{P}={\left\{(x_i,y_i)\right\}}_{i=1}^N$ positioned along the centerlines of accessible routes. Each AprilTag marker corresponds to a Path-Point. Between markers, each estimated user location $\widehat{p}=(\widehat{x},\widehat{y})$ obtained from the pedometer-based tracking is snapped to the nearest Path-Point in $\mathcal{P}$ using Euclidean distance (Equation (1)):

$$i^{\star }=arg\underset{i\in \{1,\dots ,N\}}{min}\sqrt{{(\widehat{x}-x_i)}^2+{(\widehat{y}-y_i)}^2}$$

(1)

This approach reduces the effective state space, regularizes trajectories, and ensures estimates remain within valid walkable locations. Figure 2 illustrates the concept.

Path-Points are defined by the system administrator during the initial setup phase by analyzing the floor plan, identifying key navigation paths, and positioning these points along the routes at appropriate intervals. Marker identifiers, the graph of Path-Points, and general metadata form the Base Layer, which represents the walkable geometry of the building.

An excerpt of the JSON structure encoding the Base Layer, which is sent to the client, is shown in Listing 1. At runtime, routes to user-specified destinations are computed on this graph using standard pathfinding algorithms.

Listing 1. Excerpt of the JSON structure encoding Base Layer (Path-Points, edges, and markers).

{

....

"baseLayer": {

"metadata": {

"version": "2025-09-18",

"buildingId": "B13",

....

},

"pathPoints": [

{ "id": "pp-001", "x": 30.0, "y": 30.0 },

{ "id": "pp-002", "x": 30.0, "y": 40.0 },

{ "id": "pp-003", "x": 40.0, "y": 40.0, "type": "door" },

....

],

"edges": [

{ "from": "pp-001", "to": "pp-002", "bidirectional": true },

{ "from": "pp-002", "to": "pp-003", "bidirectional": true },

....

],

"markers": [

{ "markerId": 0, "pathPointId": "pp-003" },

{ "markerId": 1, "pathPointId": "pp-010" },

....

]

...

}

3.2. Informative Layers

To extend the utility of FloorTag beyond basic position tracking, the system implements the concept of Informative Layers. Users typically navigate to destinations using semantic queries rather than coordinates, for example, locating a specific Professor’s office or finding particular services within a building. Informative Layers address this need by overlaying meaningful metadata onto the Base Layer’s spatial structure, enabling content-based search and navigation. For example, a “Professors’ Offices” layer can associate faculty names and metadata (academic discipline, contact information, office hours) with their spatial locations. In retail environments, a “Product Categories” layer could map departments like Electronics or Apparel to store aisles, enabling the routing algorithm to guide customers to relevant sections. This design enables administrators to configure application-specific layers without modifying the underlying positioning and navigation infrastructure, ensuring flexibility and scalability. The client application leverages these layers to provide enhanced search and navigation capabilities, transforming the system from a purely coordinate-based positioning tool into a context-aware, goal-oriented navigation assistant. The Informative Layers for each building floor are encoded in JSON format and transmitted to the client along with spatial data from the Base Layer.

Listing 2 illustrates an example JSON structure for the professorsOffices layer, where each Professor entry includes metadata such as contact details and subject areas. Listing 3 shows the JSON structure linking Path-Points to their associated Informative Layers.

Listing 2. Sample JSON structure for the professorsOffices Informative Layer.

{

....

"informativeLayers": {

"professorsOffices": [

{ "id": "prof-rossi",

"name": "Mario Rossi",

"phone": "+39-095-112233",

"subjects": ["Computer Science"],

....

},

{ "id": "prof-verdi",

"name": "Sara Verdi",

"phone": "+39-095-445566",

"subjects": ["Electronics"],

....

}

]

....

}

}

Listing 3. Excerpt of the JSON structure linking Path-Points to Informative Layers.

{

"version": "2025-09-18",

"buildingId": "B13",

....

"associations": [

{ "pathPointId": "pp-003",

"refs": [

{ "layerId": "professorsOffices", "entityId": "prof-rossi",

"role": "door", ....},

....

]

},

{ "pathPointId": "pp-021",

"refs": [

{ "layerId": "professorsOffices", "entityId": "prof-verdi", ....},

...

]

}

]

....

}

Figure 3 shows the floor plan of Building 13 at the University of Catania, where offices are labeled with letters (A, B, C, etc.) for simplicity. The Informative Layer “professorsOffices” is overlaid on the spatial Base Layer, indicating that Professor Mario Rossi occupies office A and Professor Sara Verdi occupies office U (other faculty members omitted for clarity). FloorTag enables users to search for a faculty member by name and receive turn-by-turn navigation from their current location to the corresponding office, without needing to know the office identifier in advance.

3.3. Server-Side Architecture

Figure 4 illustrates the overall architecture of the FloorTag system, which follows a client–server model. The server-side architecture is logically divided into the following:

A server-side front-end component which exposes an intuitive user interface to configure and manage the system. This component enables administrators to perform tasks such as (a) registration of new buildings with associated metadata (name, address, description); (b) configuring floor-specific parameters including floor number, indoor maps, and attributes; and (c) defining spatial elements such as Path-Points, AprilTag marker positions, and Informative Layer configurations. The workflow consists of several key phases, performed through a simple web-based administration tool:

• Map ingestion and calibration: The administrator uploads a high-resolution image of the building’s floor plan and defines a Cartesian coordinate system on top of the plan. A scale is established by selecting two reference points on the image (e.g., endpoints of a corridor segment with known length) and entering their physical distance, so that all coordinates are stored in meters in the resulting reference frame.
• Coordinate system definition: An arbitrary origin point is defined on the map (typically the bottom-left corner). All subsequent spatial elements are stored as 2D Cartesian coordinates $(x,y)$ relative to this origin.
• Path-Point and marker registration: Once the map is calibrated, the walkable space is discretized into Path-Points by the administrator, who interactively places points along corridor centerlines and other accessible routes in the floor reference frame, at the desired spacing (≈2 m in the prototype deployment). AprilTag markers are then associated with the Path-Points by specifying for each marker a correspondent marker identifier and selecting an existing Path-Point.
• Informative Layer Definition: To enable the goal-oriented navigation, using the calibrated map interface, specific Path-Points or spatial regions are visually selected and associated with semantic metadata by the administrator.

A server logic component that processes administrator-provided data and generates optimized JSON structures for client applications. It exposes RESTful APIs enabling clients to retrieve building metadata, spatial configurations, and Informative Layers. All communication follows standard JSON-based protocols (see Listings 1–3), ensuring platform-independent interoperability.

3.4. Client-Side Architecture

The client-side architecture is illustrated in Figure 5. The system has been implemented as an iOS smartphone application, logically structured into three primary components: Managers, View Controllers, and the User Interface.

Managers are responsible for interfacing with device hardware and handling external data sources. The Outdoor Position Manager manages outdoor-to-indoor transitions by acquiring the user’s initial location through the iOS CoreLocation framework (GPS/Wi-Fi/cellular), detecting proximity to registered buildings, and triggering appropriate notifications to view controllers. The Data Communication Manager handles client–server data exchange, downloading building configurations including Path-Points, marker metadata, and Informative Layers. The Pedometer Manager interfaces with the device’s motion coprocessor via the Core Motion framework, tracking step counts to support dead reckoning between marker detections. The Camera Manager provides continuous access to the rear camera, streaming video frames to the localization modules for near-real-time marker detection and decoding.

View Controllers manage application logic and user flow. Outdoor View Controllers present the user’s outdoor position and display the list of nearby buildings with available indoor maps. Settings View Controllers expose configurable parameters such as building search radius, step length, enabled Informative Layers, and pathfinding options. Advanced settings allow fine-tuning of AprilTag detection parameters. The Indoor View Controller serves as the central localization component. Upon building selection, it retrieves the corresponding spatial and semantic data from the server. The component continuously integrates two data streams: step counts from the Pedometer Manager and video frames from the Camera Manager. Using the AprilTag 2 library, it detects and decodes visual markers in captured frames. To keep the client-side implementation lightweight and avoid additional computational overhead, the current prototype does not perform explicit 3D pose recovery; instead, markers are used solely to correct the 2D position on the indoor map. When a marker is identified, its ID is mapped to a known $(x,y)$ position on the indoor map, and the step counter resets. Between marker detections, the controller estimates displacement from accumulated steps and the configured step length, snaps the estimated position to the nearest Path-Point, and computes routes to user-selected destinations using the spatial graph and Informative Layers. The User Interface facilitates user interaction by displaying the indoor map with the user’s real-time position and overlaying computed navigation routes. The interface updates continuously as the localization system processes new sensor data and marker detections. Figure 6 and Figure 7 show screenshots of the prototype application.

From a privacy and data protection perspective, FloorTag follows a privacy-by-design approach. All camera frames acquired by the rear camera are processed locally on the smartphone: the AprilTag detector operates directly on these frames, and no raw images or video streams are transmitted to, or stored on, the server. The only data exchanged with the server consist of building configuration data such as Path-Points-related data, marker metadata, and data about Informative Layers. The system does not perform person or face recognition, nor does it attempt to infer user identity from visual content; visual data are used exclusively to detect high-contrast fiducial markers on the floor and are discarded once processed. In real deployments, system administrators are, nonetheless, expected to comply with applicable legal and regulatory frameworks by, for example, providing clear in-app information.

4. Discussion and Results

The prototype was evaluated in Building 13 at the University of Catania, Italy. Nine AprilTag markers (IDs 0–8) were deployed on the floor following a strategic, non-uniform placement approach. Markers were positioned at key decision points, such as corridor intersections or doorways, where navigational guidance is most critical. A maximum inter-marker distance of approximately 8 m was adopted, determined empirically as a conservative threshold to ensure effective drift management: this spacing allows the system to recalibrate the user’s position by resetting accumulated errors from pedometer-based dead reckoning before they become significant, and it aligns with typical drift rates observed in smartphone-based pedestrian dead reckoning [76,77]. The system is designed for flexibility and adaptability: marker density may be increased in architecturally complex zones and reduced in long, unobstructed corridors with no points of interest. The markers belong to the pre-generated 36h11 AprilTag family, which provides 587 unique 6 × 6 bit patterns with a minimum Hamming distance of 11.

Figure 8 shows (a) the test path and (b) the building’s indoor map with marker positions. Path-Points were placed along the test route at approximately 2 m intervals to discretize the walkable space, with markers deployed at selected Path-Point locations, as shown in Figure 9.

4.1. Evaluation of FloorTag with AprilTags Only (Pedometer Disabled)

The system’s performance was first evaluated under varying lighting conditions using an iPhone SE (2nd generation) as the testing device, with the pedometer module disabled to isolate visual marker detection from dead reckoning. Initially, 3.2 × 3.2 cm printed markers were tested. Detection failures were primarily observed when (1) intense direct light reduced marker contrast and (2) users walked at normal speeds rather than slowly. While tests in a controlled environment demonstrated reliable AprilTag detection across all tested marker sizes under optimal conditions, as expected, real-world deployment introduces degrading factors such as reflections, shadows, print quality variations, and floor texture interference. Through iterative testing, it was determined that standard-quality printing with 6.0 × 6.0 cm markers provided the optimal trade-off between physical invasiveness and detection reliability, achieving near-perfect decoding rates under normal conditions. This choice is also supported by our previous peer-reviewed study on floor-deployed visual markers [75], where several widely used markers were systematically evaluated under variations in lighting, floor patterns, motion blur, and marker–camera distance using a smartphone-based acquisition setup. The results showed that AprilTag provided the most stable real-time performance across all tested conditions, and that marker sizes below approximately 5 cm degraded detection reliability, particularly under moderate or low illumination. In contrast, AprilTags in the 5–6.5 cm range consistently achieved full decoding rates, justifying the adoption of a 6.0 × 6.0 cm marker in the FloorTag system.

Rather than measuring absolute detection timing, which exhibits high variability in dynamic real-world environments, a 250 ms threshold was adopted, supported by empirical testing based on user-perceived responsiveness and established human perception studies [78], to represent near-real-time performance for indoor navigation. This threshold ensures that detection latency is minimally perceptible during navigation tasks at typical walking speeds (1–1.5 m/s). Subsequently, the percentage of markers successfully decoded within this threshold was measured across three operationally defined lighting categories, chosen to represent the range of real-world illumination conditions in the test environment. A qualitative definition was preferred over strict quantitative illuminance thresholds to ensure both reproducibility and practical applicability across different buildings, where absolute light levels may vary but relative lighting conditions remain consistently identifiable. The categories were defined as follows:

• Optimal lighting represents a best-case scenario during daytime, simulating conditions with maximum visibility. All artificial ceiling lights in the corridor were switched on, and all window curtains were fully raised to allow maximum natural light, resulting in bright and uniform illumination without significant shadows or reflections on the markers.
• Moderate lighting describes typical daytime indoor scenarios where approximately only 50% of the artificial ceiling corridor lights were activated (in an alternating pattern), and only half of the window curtains were raised. In this configuration, some visual markers were partially mildly shadowed by nearby objects, such as electronic devices (e.g., photocopiers) or recycling bins, and occasionally experienced reflections.
• Low lighting simulates a challenging, poorly lit daytime scenario, with all the corridor lights switched off and window curtains partially lowered to restrict natural light, with visual markers poorly illuminated, as they received only diffuse natural light, and partially shadowed by corridor objects.

Table 1. Percentages indicate the proportion of markers that were not decoded, decoded above 250 ms, and decoded below the 250 ms threshold, averaged across five runs (AprilTag only).

Light Conditions	Less than 250 ms	More than 250 ms	Not Decoded
Optimal Light	100%	0%	0%
Moderate Light	100%	0%	0%
Low Light	73%	13%	14%

presents qualitative results averaged across five complete traversals of the test path under each lighting condition. Under optimal and moderate lighting conditions, all markers were successfully decoded within the 250 ms threshold. Under low lighting conditions, 14% of markers failed to decode, while 13% exceeded the threshold but were eventually decoded successfully. Figure 10 shows a representative example of the navigation results: under optimal and moderate lighting conditions (a), the entire test path is tracked successfully across all markers. Under low lighting conditions (b), one marker fails to decode due to insufficient illumination and shadow interference, resulting in a gap in the tracked path.

4.2. Evaluation of FloorTag with AprilTags and Pedometer Integration

The evaluation was repeated under optimal lighting conditions with the iOS pedometer enabled. As Apple does not disclose formal accuracy specifications for this sensor, preliminary calibration tests were conducted using an iPhone SE (2nd generation) held naturally in the user’s hand, with the rear camera oriented downward toward the floor during walking. In these trials, the step count exhibited a small cumulative bias (approximately 2 steps per hundred), resulting in errors that slightly increased over longer walks. The system maps each pedometer-estimated position to the nearest Path-Point via Euclidean distance. Two categories of errors can occur during pedometer-based tracking:

• The iOS pedometer APIs introduce an intrinsic delay of a few seconds in providing step-count updates. Depending on the density of Path-Points, walking speed, and step length, the algorithm may advance past one or more intermediate Path-Points between consecutive marker detections, causing the user’s estimated position to “jump” forward along the path.
• Between two AprilTag detections, accumulated step count errors may cause the snapping algorithm to place the user at a wrong Path-Point, resulting in position estimation errors.

Figure 11 illustrates the results from one representative test. In (a), controlled conditions were maintained: approximately constant 100 cm step length and slow, deliberate walking pace. The system missed three Path-Points (primarily during the first few seconds, corresponding to the pedometer’s initialization period) but mapped the user to the correct Path-Point thereafter. In (b), results obtained under uncontrolled conditions are shown, characterized by a natural, moderate walking speed without enforcing step length constraints. In this case, three Path-Points were missed and occasional incorrect position assignments occurred due to step length variability and accumulated drift.

Table 2. Percentage of missed Path-Points and wrong positions for controlled and uncontrolled walks (with Pedometer Integration).

Metric	Controlled Walks	Uncontrolled Walks
Missed Path-Points (%)	21%	19.3%
Wrong positions (%)	2.2%	14.2%

summarizes results from five complete runs under each condition. The rate of missed Path-Points remains similar between controlled (21%) and uncontrolled (19.3%) walks, as missing is primarily determined by the pedometer’s update latency rather than step accuracy. However, incorrect position assignments increase significantly from 2.2% (controlled) to 14.2% (uncontrolled), reflecting the impact of step length and walking pace variability on cumulative position error.

It is important to note that FloorTag is evaluated at the Path-Point level, reflecting its design as a discretized navigation system. Along the test route, Path-Points are placed at approximately 2 m spacing along the corridor, and the user’s estimated position is always snapped to the nearest Path-Point. Under this model, localization accuracy is naturally expressed in terms of node-level correctness: a “correct” assignment indicates that the user lies within the validity region of the corresponding Path-Point, whereas a “wrong” position corresponds to snapping to a different node. The percentages reported in Table 2, therefore, quantify the probability of selecting the correct node at a spatial granularity of about 2 m, rather than a continuous Euclidean error in meters. Given the small sample size and the proof-of-concept nature of the study, formal statistical significance testing was not performed; instead, the objective was to characterize the typical behavior of the prototype under realistic usage conditions.

5. Conclusions and Future Work

This work presents FloorTag, a hybrid IPS that combines floor-deployed AprilTag visual markers with pedometer-based dead reckoning and enables semantic, goal-oriented navigation. The system addresses common limitations of existing IPS solutions discussed in Section 2 by minimizing infrastructure requirements, deployment complexity, and associated costs while maintaining usability for non-technical administrators. The system makes three primary contributions: (1) it leverages floor-mounted markers for passive, unobtrusive detection (rear cameras naturally point downward during walking, and floor surfaces provide predictable backgrounds with stable camera-to-marker distances); this design choice proves effective compared to wall-mounted or active-scanning alternatives. (2) It introduces Path-Points to discretize walkable space and constrain position estimates to valid locations. (3) It employs Informative Layers to enable semantic queries rather than coordinate-based navigation. Experimental evaluation of an iOS prototype deployed in Building 13 at the University of Catania demonstrates near-real-time performance (sub-250 ms marker detection) and effective 2-m spatial granularity along discretized walkable routes. Under optimal and moderate lighting conditions, the system achieves 100% marker detection rates with 6.0 × 6.0 cm printed markers. Performance degrades moderately under low-light conditions (73% detection), confirming robustness across realistic operating environments. Integration with pedometer-based dead reckoning provides continuous tracking between markers, with controlled walking achieving 2.2% incorrect position assignments and natural walking reaching 14.2%. The evaluation methodology prioritizes real-world applicability over laboratory precision. A qualitative lighting categorization (optimal, moderate, low) is adopted rather than photometric measurements, reflecting how building administrators would assess deployment environments visually. Similarly, the 250 ms responsiveness threshold captures user-perceived performance rather than absolute timing metrics. Indoor environments exhibit complex, dynamic lighting (natural variation, artificial sources, occupant shadows, and reflective surfaces), making precise measurements less representative than operational categories. This approach aligns with FloorTag’s core design philosophy: practical deployment by non-specialist personnel in diverse real-world settings. In addition to performance and deployment aspects, FloorTag’s client-side design limits the circulation of potentially sensitive data, as raw camera frames remain on the device and are used solely for on-device marker detection. This privacy-by-design choice aligns with contemporary data protection expectations and facilitates the adoption of FloorTag in privacy-sensitive environments such as hospitals, public buildings, and educational institutions.

Building upon these findings, several directions for future research and development have been identified to enhance the system’s robustness, scalability, and operational efficiency. A first line of investigation concerns cross-platform evaluation and marker optimization. The current prototype was assessed using a single device (iPhone SE, 2nd generation), deliberately chosen to demonstrate feasibility on older, computationally limited hardware. Future work will extend this evaluation to a broader range of recent iOS and Android devices to quantify variability in detection speed, robustness under challenging lighting conditions, and overall computational performance. This cross-device analysis will also clarify the extent to which advances in camera sensors and mobile processors enable the use of smaller visual markers, potentially reducing the current $6.0\times 6.0$ cm size while preserving near-real-time responsiveness, an important requirement for deployments in aesthetically sensitive environments. Moreover, a promising direction is to exploit the full six degree-of-freedom pose of the camera with respect to each AprilTag marker. Integrating such orientation estimates into the dead-reckoning pipeline would allow the explicit correction of heading drift at each marker detection, improving the robustness and accuracy of the system. A second research direction involves assessing scalability in more complex architectural scenarios. The present study was conducted in a single-floor building with relatively linear corridors, representing a controlled and structurally simple layout. Evaluations in larger, multifloor structures, such as hospitals, museums, or shopping centers, will allow examination of FloorTag’s performance under denser pedestrian traffic, diverse lighting conditions, and frequent occlusions caused by moving occupants. These experiments will offer insight into optimal marker placement strategies, appropriate Path-Point density, and refinements to inertial dead reckoning required to handle longer intervals without visual updates. In addition, while the current evaluation adopts discretized metrics consistent with FloorTag’s topological nature, it does not provide a continuous spatial error distribution. Future developments will enhance the validation framework by integrating ground-truth tracking, thereby supporting rigorous numerical comparison with conventional IPS benchmarks. Further improvements may arise from advanced sensor fusion, supported by FloorTag’s modular architecture. Future research will explore integrating BLE beacons to provide coarse, background localization that initializes the user’s floor or building wing before visual tracking begins. Additional corrective layers, such as Wi-Fi fingerprinting or magnetic field signatures, could offer increased robustness in situations where markers are temporarily occluded or insufficiently illuminated. Moreover, coupling smartphone sensors with those of a smartwatch may improve dead-reckoning accuracy through sensor redundancy and complementary motion patterns. Energy consumption should also be taken into account. Keeping the camera and decoding pipeline continuously active may affect battery life, potentially limiting long-duration usage. Investigating adaptive frame-rate policies, such as reducing processing frequency when users remain stationary, or adjusting processing based on marker density and navigation context could provide a more effective balance between responsiveness and energy preservation. From an engineering perspective, the system can evolve toward more integrated or temporary infrastructure solutions. Permanent installations could embed fiducial markers directly into floor tiles during manufacturing, while temporary deployments, such as trade fairs or exhibitions, could utilize removable carpets with pre-printed markers. Such solutions, combined with a server-side configuration interface, would enable rapid deployment and flexible reconfiguration for different spatial layouts, and could be employed to guide visitors through stands and points of interest during temporary events, effectively providing a plug-and-play navigation system. Finally, future work will explore accessibility-oriented enhancements, adapting the user interface for visually impaired individuals. Because FloorTag operates without requiring users to deliberately aim the camera at wall-mounted targets, it is compatible with screen readers and can be extended with haptic feedback to support turn-by-turn indoor guidance.