In complex indoor environments, reliable positioning without Global Navigation Satellite Systems (GNSS) coverage poses a significant challenge, particularly for emergency responders operating in areas without installed building technology (e.g. Wi-Fi, BLE, UWB). This paper presents a comparative study of two infrastructure-free indoor localization approaches, both implemented on standard smartphones: inertial Pedestrian Dead Reckoning (PDR) and vision-based Simultaneous Localization and Mapping (SLAM). Each prototype is evaluated based on several criteria, including positioning accuracy, environmental dependencies, hardware requirements, and usability. While the PDR system ofers robust, low-power tracking independent of visual conditions, the SLAM-based system achieves higher precision and supports augmented reality (AR) navigation under favorable lighting. A structured comparison is presented to highlight the strengths, limitations, and potential use cases for each approach, based on practical evaluations. The findings ofer actionable insights for the future development of indoor positioning systems tailored for first responders in GNSS-denied environments.
The positioning of first responders in complex indoor environments poses a particular challenge due
to the insuficient strength of GNSS signals in such environments. Furthermore, the buildings in
question are often unfamiliar upon their initial entry, and relying on preinstalled building infrastructure,
such as Wi-Fi access points or Bluetooth beacons, is not feasible. Consequently, the Federal Agency
for Cartography and Geodesy (BKG) is engaged in research to develop systems that will empower
emergency services to determine their location in areas lacking GNSS coverage. Instead of existing
location-based infrastructure, compact inertial navigation systems (INS) are implemented, which utilize
integrated sensors - such as accelerometer, gyroscope, and magnetic field sensors — to detect the
movement of the emergency responder within the room and derive their current position [
This paper presents a comparative study of two infrastructure-independent approaches to indoor localization, both implemented as smartphone-based prototypes. The key contributions of this work are as follows: https://www.bkg.bund.de/deniedAreas/ (P. Fiedler)
CEUR
ceur-ws.org • Implementation of two distinct prototypes for infrastructure-free indoor positioning: one based on inertial Pedestrian Dead Reckoning (PDR), the other on visual SLAM. • Qualitative comparison of both approaches on positioning accuracy, drift behavior, robustness, sensor requirements, and environmental dependencies. • Identification of practical strengths and limitations of each approach based on real-world usage scenarios and developer observations. • Presentation of a structured comparison table to support decision-making in future applications, such as emergency response, indoor navigation, or infrastructure-free localization.
The goal is to derive practical insights into the applicability of PDR and SLAM-based localization methods under constrained conditions, such as those encountered by first responders in unfamiliar indoor environments without prior mapping or network availability.
This comparison underscores the trade-ofs between diferent sensing technologies. Inertial sensors such as accelerometers and gyroscopes are universally available but sufer from cumulative drift. Vision-based approaches ofer high spatial accuracy, but are sensitive to lighting and require significant computational resources.
One possibility that has been extensively discussed in the literature as a way of eliminating the
aforementioned disadvantages of individual sensors is to combine diferent sensors within a single
application [
Kalman filters, particle filters, and machine learning-based models are commonly used to fuse sensor
data, compensate for drift, and estimate user orientation and movement [
This paper builds on these foundations by evaluating two sensor fusion-based prototypes that follow fundamentally diferent approaches: one primarily inertial (PDR) and one vision-based (SLAM), both implemented on commodity smartphones.
In order to enable optimal indoor positioning that is independent of GNSS, the Internet, or known infrastructures such as Bluetooth beacons, two prototypes were developed based on diferent approaches: sensor-based PDR and visual SLAM. The two systems were developed for Android smartphones and are designed to be as universally applicable as possible in all buildings without external infrastructure. While the PDR prototype focuses solely on real-time positioning (Figure 2 A), the Visual SLAM prototype additionally provides navigation capabilities, ofering a AR-based interface with route guidance for the user (Figure 2 B and C).
Despite the challenges posed by the use of GNSS within structures due to the attenuation of signal strength, it remains a viable method to determine the initial starting point, if the user starts the application outside of the building. The system performs a variety of checks to assess the quality of the GNSS signal for positioning purposes, including the Carrier-to-Noise Ratio (CNR), the number of visible satellites, and the reported GNSS accuracy. If the user initiates the process within the building or encounters a disturbance in the GNSS signal, manual entry of a starting point is allowed. To determine the initial heading, the system uses a magnetometer to obtain an absolute orientation relative to the magnetic north. Given the potential for interference signals in this environment, GNSS direction or manual alignment to the north may be considered as alternatives for initial heading determination. (A) (B) (C)
Once the initial pose (position and heading) is established, step detection is performed using
accelerometer data. The step length () is dynamically estimated using the Weinberg model [
() = ⋅ √4 peak() − min()
The heading () is continuously updated using fused gyroscope and magnetometer data. Using these values, the local coordinate diferences for each step can be calculated starting from the initial position using formulas (2) and (3).
These diferences are then converted into global coordinates: longitude = 111,320 ⋅ cos ( latitude⋅ ) 180 The system also estimates the user’s vertical position (floor level) by comparing the height delta to the starting point using the barometer. For this computation, a constant floor height of 3 m can be assumed. Δ = () ∗ sin(()) Δ = () ∗ cos(()) latitude =
Δ 111,229 Δ (1) (2) (3) (4) (5)
To mitigate drift and cumulative positioning errors, several correction mechanisms are employed. The system continuously scans for usable GNSS signals. In some indoor environments, signal quality is suficient to apply GNSS-based corrections. Additionally, users may temporarily leave the building, providing an opportunity for re-initialization via satellite data. Magnetometer readings are used to compensate gyroscope drift. Various filters determine which sensor should be given more weight based on magnetic field strength.
The smartphone interface shows the user’s current position and orientation on an OpenStreetMap background, along with a GNSS accuracy circle. The minimum acceptable GNSS accuracy threshold can be adjusted using the slider at the bottom of the screen. Users also have the option to track and save their traveled path as GeoJSON. The operations center receives this information at 1 Hz intervals via WiFi Websockets, enabling real-time monitoring of the emergency responder’s current position.
In contrast to sensor-based approaches, the Visual SLAM system utilizes a QR-Code scan to initialize the user’s pose. The QR-Code is generally positioned at the building entrance, encoding both the absolute position and the orientation of the smartphone at the moment of scanning. This approach ensures a precise and unambiguous starting pose, which is essential for global map alignment and consistent localization in the building.
After determining the initial pose, the system utilizes Google’s augmented reality SDK ARCore to detect and track visual feature points from the camera stream. These features, including corners, edges, and texture regions, are extracted at a frequency of 60 Hz and tracked over time to estimate the smartphone’s motion through space. When the same feature is observed across multiple frames from diferent viewpoints, its 3D position is triangulated.
To improve motion estimation, ARCore fuses data from the device’s gyroscope, which helps infer rotation and movement direction, especially during rapid motion or in low-feature environments. The estimated trajectory is then aligned with a preloaded three-dimensional reference map of the building to provide global positioning. The 3D map can be derived from evacuation and emergency plans, which are available to the authorities for all public buildings and must be continually updated, making them ideal for SLAM-based navigation and matching. In the current implementation, this conversion is performed manually in Unity3D, a process that can take up to 30 minutes. Subsequently, a topological representation (NavMesh Surface) is automatically generated using Unity’s AI Navigation tool.
The system is supported by additional sensors, including accelerometer for motion detection and a barometer for altitude detection, similar to prototype A. When the system revisits a previously recorded location, a process known as loop closing is initiated to address accumulated errors resulting from sensor noise or drift. This objective is accomplished by identifying and re-matching features that were recorded during the initial stages of the SLAM process.
Beyond mere positioning, the system provides comprehensive indoor navigation capabilities. According to the present position and the user-defined destination, the system continuously performs computations using an A* algorithm to ascertain the shortest path. In this process, multiple environmental constraints (walls, rooms, impassable areas) are taken into account, which are defined in the topological map.
Navigation instructions are displayed in Augmented Reality (AR), overlaid directly onto the smartphone camera view. Users have the option to select diferent routing preferences, such as avoiding stairs or choosing between two visualization modes (arrow-based overlay vs. floating path line). A minimap is displayed alongside the AR view to facilitate user orientation and provide an overview of the current route within the building structure.
Both approaches have been developed as prototypes and tested under realistic indoor conditions. This has provided valuable insights into the respective strengths and weaknesses of the two approaches, which will help to define the potential use cases and deployment scenarios for sensor- and vision-based indoor positioning.
The evaluation was conducted across diverse environments, including residential buildings, ofice complexes, and large-scale structures such as football stadiums. The predefined multi-level trajectories, varying in length from approximately 50 meters to several hundred meters, were traversed using both systems. At each step, the estimated position was recorded. These recorded positions were compared against the reference trajectory, and the mean positional accuracy was calculated. For a fair comparison, both systems were carried in a normal usage position, similar to how a smartphone is typically held during navigation tasks (e.g., in front of the user or in hand at waist level). The primary advantage of the PDR-based system is its universality, which renders it applicable in the absence of any prior knowledge of the building environment. The system operates independently of infrastructure and can be deployed in nearly any indoor setting. Additionally, the smartphone does not need to be held in the user’s hand: Due to its robustness against poor lighting or visual obstructions, the placement of the device in a pocket or on the body of the first responder is suficient.
The PDR algorithm’s low computational requirements and energy consumption make it suitable for execution on low-power microcontrollers (e.g. ESP32) equipped with low-cost inertial and barometric sensors. The BKG has already implemented several such prototypes.
However, PDR is susceptible to cumulative drift over time, particularly in scenarios where reliable GNSS or magnetometer data is not available. In such cases where only the gyroscope is utilized for heading estimation, the resulting drift is approximately 2% of the distance. The system’s performance is contingent upon the precise initialization of both position and heading. Without a fix, even small initial errors can lead to significant positional errors as the user moves, leading to a mean accuracy tested of about 5m.
The visual SLAM approach addresses many of the limitations of PDR through continuous visual relocalization and loop closing, significantly reducing long-term drift. Under optimal visual conditions, characterized by a light intensity greater than 10 lux, the camera-based system achieves high accuracy of around 60 cm and is expected to be less sensitive to initialization errors, as the initial pose is determined via a fixed QR code position rather than GNSS.
However, the system is sensitive to low-light conditions and motion blur, which can impair feature tracking and stability. Moreover, measurements of processing power utilization and battery consumption indicate that visual SLAM demands approximately 3.3 times greater computational power and about 4.5 times higher energy consumption. For optimal performance, the smartphone must be actively held in the user’s hand in a way that ensures an unobstructed camera view.
The following table 2 summarizes the key characteristics of both systems in direct comparison:
A comparison of the two prototypes reveals the distinct strengths of PDR and visual SLAM in terms of operational applicability. Depending on the specific requirements of the respective operation, each approach lends itself to distinct use cases.
The PDR-based system is particularly well suited for spontaneous operations where minimal setup time is available and first responders may not be able to hold a smartphone in their hands. In many of these scenarios, room-level accuracy (approximately 5 m) is suficient to track and coordinate personnel from a command center. The following list contains some of the more common use cases: • Tracking and coordination of first responders during dynamic or ad hoc operations • Operations in darkness or low-light conditions (e.g., night-time or smoke-filled environments) • Situations where responders require hands-free operation
In contrast, the camera-based SLAM approach is better suited for pre-planned deployments in public or complex buildings, where suficient preparation time is available to generate a 3D map from oficial building evacuation plans. The system’s enhanced accuracy and AR navigation capabilities support the following scenarios: • Post-analysis and documentation of tactical training exercises • Pre-mission preparation in structured public environments (i.e., train stations, malls, government buildings) • Real-time indoor navigation and tracking during static or longer-term operations
In order to eliminate the disadvantages of both systems as much as possible, research could be conducted in the future on a hybrid system that combines the advantages of both approaches. In conditions that allow for optimal visual perception, the potential system could leverage visual SLAM for high-precision positioning and use it to correct accumulated drift inherent in inertial-based tracking. In scenarios where visual tracking becomes unreliable due to occlusions, darkness, or motion blur, the PDR system could seamlessly take over, maintaining continuous position estimations in the absence of visual input.
This fusion would enable more robust and adaptive indoor localization, particularly in complex environments with rapidly changing conditions. Developing such a hybrid system would require careful synchronization of sensor data and intelligent switching or fusion strategies, but holds significant potential for improving overall reliability, particularly in mission-critical scenarios.
This paper presented a comparative analysis of two infrastructure-independent approaches to smartphone-based indoor localization: one based on inertial PDR and the other on visual SLAM. While the PDR prototype demonstrated robustness and low dependency on environmental factors, the visual SLAM system ofered significantly higher accuracy under favorable conditions.
The comparison indicated that each system is best suited to diferent operational contexts, ranging from hands-free emergency deployments to high-precision navigation in structured environments. A hybrid integration of both approaches could further improve adaptability and reliability, especially in dynamic or degraded settings.
Future work will focus on further improving both systems and exploring a hybrid solution that
combines their respective strengths. The ongoing development includes a lightweight, ESP32-based
version of the PDR system, investigations into alternative positioning technologies such as DAB+,
and integration of LoRa for remote transmission of positional data. The PDR algorithm is undergoing
enhancements that include improved dynamic step length estimation based on [
During the preparation of this work, the author(s) used DeepL Write in order to: Grammar and spelling check. After using these tool, the author reviewed and edited the content as needed and takes full responsibility for the publication’s content.