Indoor Multimedia Geolocation

• Indoor multimedia geolocation is a multidisciplinary approach that fuses radio, vision, audio, and inertial signals to determine indoor positions and semantic areas when GPS is unreliable.
• Techniques involve radio fingerprinting, visual landmark retrieval, echo-based SLAM, and crowd-sourced floor-plan construction, often achieving sub-meter accuracy in controlled deployments.
• Recent advances leverage multimodal fusion, machine learning, and hybrid sensor strategies to enable robust, scalable, and privacy-preserving indoor localization and semantic mapping.

Indoor multimedia geolocation denotes the estimation of an indoor building, floor, coordinate, trajectory, or semantic area for multimedia recordings, mobile devices, or users in environments where GPS is unreliable or unavailable. In the literature, the problem is formulated in several equivalent ways: as indoor positioning from radio, inertial, or mmWave measurements; as floor-plan and area-of-interest construction from crowdsensed smartphone traces; as visual landmark retrieval over panoramic or ground-texture databases; as echo-based SLAM using smartphone audio; and as multimodal fusion in which camera, radio, and inertial signals jointly constrain location (Youssef et al., 2012, Hu, 2020, Luo et al., 2022, Papaioannou et al., 2023). A distinctive feature of the topic is that indoor location is often not only geometric but also semantic: a system may infer that a capture belongs to a corridor, office, stairs, elevator, room, or country-specific indoor context rather than merely outputting a point estimate (Youssef et al., 2012, Aftab et al., 18 Dec 2025).

1. Problem formulation and representational scope

Indoor geolocation is motivated by the failure of satellite positioning in enclosed spaces and by the operational demands of airports, hospitals, museums, malls, offices, and public buildings. Airport-oriented work emphasizes that indoor localization should work where GPS fails, be usable through familiar devices, be inexpensive to deploy, scale to many users, and preserve privacy (Santos-González et al., 2022). Comparable motivations appear in Android navigation systems for large buildings, where wall-mounted maps and signs remain the primary reference indoors, and in ubiquitous indoor positioning architectures intended to support seamless direction finding between indoor locations, first responders’ safety, and richer location-aware services (Ramani et al., 2014, Youssef et al., 2012).

The representational target is broader than a Cartesian coordinate. A crowd-sourced indoor positioning architecture explicitly separates a Floor Plan Construction Module, a Location Information Database Builder, a Location Determination Engine, and a User Interface Module, while treating camera images and sound signals as first-class data streams alongside accelerometer, compass / magnetometer, gyroscope, WiFi RSSI and AP MACs, GSM RSSI and nearby towers, and GPS when available (Youssef et al., 2012). In that formulation, indoor multimedia geolocation becomes possible when a system can associate an image, audio clip, or sensor-rich mobile capture with a building, a floor, and an area of interest such as a corridor, office, meeting room, elevator, or stairs (Youssef et al., 2012).

A parallel line of work reformulates indoor localization directly as a multimedia retrieval problem. In GPU-accelerated panoramic localization, each physical landmark is modeled as an indexed omnidirectional image feature, and user position is inferred by retrieving the most similar visual landmark from a database and mapping it back to floor-plan coordinates (Hu, 2020). This retrieval-centric view places indoor geolocation squarely within multimedia indexing, descriptor design, and candidate aggregation rather than only within classical positioning.

2. Spatial substrates: floor plans, graphs, and semantic zones

A central technical requirement is a spatial substrate on which multimedia observations can be geotagged. One approach is automatic floor-plan construction from crowd-sourced traces. In the ubiquitous IPS architecture, traces are aligned using reference points such as the last GPS point before entering the building, AP locations estimated as positions where AP signal is strongest, intersections between traces, images and sound signatures, and known landmarks such as stairs and elevators (Youssef et al., 2012). Relative localization then follows a dead-reckoning model,

$$X_k = X_{k-1} + S \cdot \cos(\theta), \qquad Y_k = Y_{k-1} + S \cdot \sin(\theta),$$

with heading from the magnetometer and distance from the accelerometer (Youssef et al., 2012).

Because direct double integration of acceleration is too noisy and the paper reports that position error can grow cubically in time and reach about 100 meters after 1 minute, the prototype replaces numerical integration with Zero Velocity Update ideas and an FSM-based step detector (Youssef et al., 2012). The failed baseline based on a local variance threshold had an average step-count error of 51.9%. The proposed detector used the states State_Stationary, State_Walking, State_Peek_Pos, State_Peek_Neg, and the noise-tolerant states State_P2 and State_W2, with steps detected by the sequence $$(\text{State Stationary}, \text{State Walking}, \text{State Peek Pos}, \text{State Peek Neg}, \text{State Stationary})$$. Reported trace-construction performance was 2.3% average step-count error and less than 0.9 m maximum displacement error over a 7-step scenario (Youssef et al., 2012).

The same system turns geometry into semantics through block-based classification. With features including average time spent in a block, average and variance of acceleration, correlation between accelerometer axes, average number of turns per trace, and average WiFi and GSM RSSI, a C4.5 decision tree with bootstrap aggregation (bagging) classified office, corridor, elevator, and stairs regions. The best result occurred at a block size of 0.7 m, with classification accuracy: 90%; elevators and corridors were detected with zero false negatives, and elevators and offices had zero false positives (Youssef et al., 2012).

Other systems encode indoor structure more explicitly as a graph. PFML discretizes the floor plan as

$G=(\Upsilon, I, \xi),$

where each valid walkable node carries Cartesian coordinates and a room identifier, with nodes separated by 0.25 m in rooms and corridors (Carrera et al., 2018). More recently, work on crowdsourced radio maps argues that such maps encode movement regularities akin to the constraints imposed by floor plans, and therefore can substitute for floor plans when explicit building geometry is unavailable or outdated (Yi et al., 2023). This suggests that indoor multimedia geolocation increasingly relies on weak structural priors learned from traces as much as on manually curated floor plans.

3. Radio and inertial localization substrates

Radio fingerprinting remains a foundational indoor geolocation technique. Android-based systems for indoor navigation on Google Maps collect RSSI values from Wi‑Fi beacons at known locations in an offline radio-map phase, then localize a user online by nearest-neighbor matching in signal space (Ramani et al., 2014). The same paper contrasts this with propagation-based localization and notes that such methods are prone to large errors indoors due to multipath and shadowing, with errors that can be up to 50% (Ramani et al., 2014). WiFiPos extends this line by preprocessing RSSI into multiple fingerprint variants—Maximum, Minimum, Mode, Average, Quartile Mode, Quartile Average, Mean Value, and Quartile Mean Value—and matching live vectors by Euclidean distance. In a 36 $m^2$ library with 36 sampling points and more than 3600 samples, the tool reported that it can determine the position of the MD with relative error up to 4.2 m in 95% of cases (Toloza et al., 2014).

The cost and fragility of dense Wi‑Fi survey have led to alternatives and hybrids. Ambient FM localization uses RSSI-based fingerprinting over broadcast FM stations rather than indoor Wi‑Fi infrastructure. In a room-level office space (12 × 6 m), the best FM result, achieved by SVM, had a median error of 0.6 m, and 40% of locations are recognized exactly. In a floor-level corridor area (50 × 25 m), the best FM result came from kNN, with 52% of locations correctly recognized and error below 24 m at 90% probability. FM was also substantially more power efficient than Wi‑Fi, providing 2.6 to 5.5 times longer battery life than Wi‑Fi, with 27.9 h battery life using 3 beacons on a Samsung Omnia2 (Popleteev et al., 2013). The same study is explicit that FM performs best in small confined areas like rooms, while in larger floor-scale environments the error increases substantially (Popleteev et al., 2013).

PFML addresses survey reduction by learning only room-level landmarks rather than dense coordinate fingerprints. It fuses Wi‑Fi RSSI, magnetic field, and floor-plan information in an enhanced particle filter, using KStar for room recognition and a hybrid NLR-LDPL model for Wi‑Fi ranging. With 2500 particles, PFML achieved average localization error: 1.55 m, standard deviation: 0.73 m, and 90% accuracy: 2.8 m, while reducing offline effort to 49 minutes total compared with 460 minutes for KNN fingerprinting (Carrera et al., 2018).

In security-sensitive airport environments, a different hybrid architecture uses QR codes, a foot-mounted IMU, and a smartphone. A QR code encodes building, floor, and position to initialize location; a foot-mounted IMU streams accelerometer, gyroscope, and magnetometer data via BLE; and the Android app converts units, applies the Madgwick filter to estimate orientation in quaternion form, and updates position by step detection and step length estimation. The step-length model is

$l = h \cdot k,$

with $k = 0.415$ for men and $k = 0.413$ for women (Santos-González et al., 2022). Reported experiments found that the Madgwick filter gave the best match between real and estimated position (Santos-González et al., 2022).

4. Vision-based and camera-assisted methods

Visual indoor geolocation spans pure image retrieval, close-range texture localization, and camera-assisted calibration for non-visual deployment. Panoramic retrieval systems model each landmark with a rotation-invariant omnidirectional descriptor obtained by transforming the omnidirectional image into a one-dimensional omnidirectional vector and using its FFT magnitude as the descriptor (Hu, 2020). Retrieval is organized with three nested levels of parallelism: database partitioning into subspaces such as floors, multi-frame query expansion, and thread-level GPU distance computation. In the reported setup, the system used $N = 15$ top candidates per frame, a query window of $M = 11$ frames, and $P = 5$ database paths; coordinates were discretized at 30 cm resolution, the tolerant aggregation radius was 3 m, and real-time performance of about 14 fps was reported (Hu, 2020).

A much finer-grained visual method is Micro-GPS, which uses a downward-facing camera to capture distinctive local keypoints in ground textures such as carpet, tile, concrete, wood, and asphalt (Zhang et al., 2017). The system uses the SIFT DoG keypoint detector, the SIFT gradient-orientation histogram descriptor, 50 randomly selected keypoints per map image, descriptor compression to 8 or 16 PCA dimensions, 10 scale buckets, and FLANN indexing. Query localization proceeds through nearest-neighbor matching, voting for the image origin on a 50 × 50 pixels grid, and local verification with RANSAC (Zhang et al., 2017). Success rates with 16D descriptors exceeded 90% on most datasets, including 99.92% on carpet and 98.40% on tiles, while wood was the hardest case at 77.49% (Zhang et al., 2017).

Vision can also serve as a temporary calibration modality for radio localization. A 2025 framework uses an overhead camera calibrated once with ArUco markers and Perspective-n-Point (PnP), then tracks a person carrying an RSS collector through a YOLOv8 + SORT pipeline (Bilge et al., 26 Sep 2025). The lower midpoint of the bounding box is projected to the floor plane to create synchronized $$(\text{State Stationary}, \text{State Walking}, \text{State Peek Pos}, \text{State Peek Neg}, \text{State Stationary})$$0 labels. Temporal alignment is enforced with NTP, bounding inter-device mismatch to less than 10 ms. For the baseline configuration—camera height 3 m, vertical FoV 5 m, tag misplacement 10 cm—the reported component errors were $$(\text{State Stationary}, \text{State Walking}, \text{State Peek Pos}, \text{State Peek Neg}, \text{State Stationary})$$1 m, $$(\text{State Stationary}, \text{State Walking}, \text{State Peek Pos}, \text{State Peek Neg}, \text{State Stationary})$$2 m, $$(\text{State Stationary}, \text{State Walking}, \text{State Peek Pos}, \text{State Peek Neg}, \text{State Stationary})$$3 m, and $$(\text{State Stationary}, \text{State Walking}, \text{State Peek Pos}, \text{State Peek Neg}, \text{State Stationary})$$4 m, yielding $$(\text{State Stationary}, \text{State Walking}, \text{State Peek Pos}, \text{State Peek Neg}, \text{State Stationary})$$5 m (Bilge et al., 26 Sep 2025). On the resulting dataset, a lightweight 1-D CNN achieved 0.094–0.144 m mean error across recordings, whereas kNN achieved about 1.86–2.38 m (Bilge et al., 26 Sep 2025).

A recurrent theme is that vision need not be the runtime sensing modality. In panoramic retrieval it is the runtime modality; in ground-texture localization it is both the map and query signal; in vision-assisted Wi‑Fi calibration it is explicitly discarded after label generation, so that deployment uses RSS only (Hu, 2020, Zhang et al., 2017, Bilge et al., 26 Sep 2025).

5. Audio, acoustics, and reflected-wave spatial inference

Audio enters geolocation in two distinct ways: as semantic content and as an active probing signal. Audio content based geotagging in multimedia uses semi-NMF to decompose an audio recording $$(\text{State Stationary}, \text{State Walking}, \text{State Peek Pos}, \text{State Peek Neg}, \text{State Stationary})$$6 with respect to sound-class bases $$(\text{State Stationary}, \text{State Walking}, \text{State Peek Pos}, \text{State Peek Neg}, \text{State Stationary})$$7,

$$(\text{State Stationary}, \text{State Walking}, \text{State Peek Pos}, \text{State Peek Neg}, \text{State Stationary})$$8

and then uses the composition of sound events such as car horn, children playing, dog barking, air conditioner, drilling, engine idling, gun shot, jackhammer, siren, and street music for city-level classification (Kumar et al., 2016). The paper explicitly states that extension to indoor geolocation is conceptual rather than demonstrated, but it establishes the broader principle that sound composition can function as geographic evidence (Kumar et al., 2016).

ELF-SLAM makes that principle indoor and metric. A smartphone emits a near-inaudible logarithmic chirp in the 15–20 kHz band, with 10 ms chirp duration and 44.1 kHz sample rate, records for 100 ms, discards the first 10 ms direct path and the next 1 ms first reflection from the human body, and uses the remaining 50 ms as the echo trace (Luo et al., 2022). A ResNet-18 encoder with a 3-layer projection head produces a 128-dimensional Echoic Location Feature (ELF) from a $$(\text{State Stationary}, \text{State Walking}, \text{State Peek Pos}, \text{State Peek Neg}, \text{State Stationary})$$9 spectrogram. Loop closures are detected by echo sequence similarity and inserted into a graph-SLAM backend implemented with g2o (Luo et al., 2022). Reported median localization errors on reconstructed trajectories were 0.1 m in a living room, 0.53 m in an office, and 0.4 m in a shopping mall, outperforming Wi‑Fi and geomagnetic SLAM baselines (Luo et al., 2022).

Wave-based geometry provides another route. mmReality uses a single fixed transmitter and a mobile receiver in the 60 GHz mmWave band, with one phased array and one RF chain at each end (Sun et al., 2022). Beam-pair probing recovers element-level observations; spatial smoothing and 2D MUSIC estimate AoAs and AoDs; multi-carrier ranging estimates path lengths; and virtual-transmitter geometry reconstructs walls from first-order reflections (Sun et al., 2022). In an irregular corridor deployment, the paper reports AoA/AoD estimation errors below $G=(\Upsilon, I, \xi),$0 at 90% probability, path length errors of 0.15 m for LoS and 0.25 m for NLoS, 0.42 m average measurement-point localization error, and 1.0 m average error for AoA-spectrum-based device localization with single AP (Sun et al., 2022).

A related but simulation-only framework, ILDARS, localizes a source from a direct path, a wall reflection, and the time difference between them. Its self-calibration stage clusters measurements by reflecting wall using Inversion or Gnomonic Projection, estimates wall direction from pairwise nested cross products, and compares All Pairs, Disjoint Pairs, and Overlapping Pairs averaging strategies (Mohanty et al., 2023). The paper concludes that Inversion + All Pairs + Narrowest Cluster + Closest Lines gives the best median performance, while Closest Lines Extended improves mean performance by using multiple walls (Mohanty et al., 2023). This suggests that reflected-signal geolocation can be treated as a geometry reconstruction problem rather than only as a nuisance-multipath problem.

6. Multimodal fusion, privacy, and application regimes

The strongest recent results often arise from explicit multimodal fusion. RAVEL (Radio And Vision Enhanced Localization) combines anonymous camera detections from a stationary calibrated camera with user-specific WiFi RSS from a carried device (Papaioannou et al., 2023). Visual detections are linked into short tracklets by motion smoothness; tracklet trees are expanded with synthetic detections and empty detections; and trajectory hypotheses are scored by

$G=(\Upsilon, I, \xi),$1

where the radio term is based on the log-normal shadowing model and the visual term uses a Kalman-filter-like motion predictor (Papaioannou et al., 2023). In a three-story museum building with a camera mounted 10 m above the ground over an 11 m × 12 m area, crowding of 8 to 20 people, and 4 people carrying WiFi-enabled smartphones, RAVEL achieved median offline error: 0.56 m and 90th percentile error: 1 m; up to 70% of windows had zero overlap error compared with 50% for vision-only tracking (Papaioannou et al., 2023).

A different fusion strategy uses radio not for identity-bearing trajectory selection but for probabilistic coarse-to-fine narrowing. The joint visual and wireless feature approach first computes SVM-based Wi‑Fi likelihoods over reference points and areas, selects the top $G=(\Upsilon, I, \xi),$2 candidate areas, and then applies a GoogLeNet backbone with four-direction LSTMs to regress position from images only within the Wi‑Fi-selected region (Wang et al., 2020). In a 4000 square meters office corridor environment, the method achieved median localization error: 0.62 m, 57% of cases within 1 m, and average total running time: 0.18 s (Wang et al., 2020). The coarse area prior is therefore used to reduce both ambiguity and search complexity.

Privacy and traceability have become explicit design variables rather than afterthoughts. In the airport QR+IMU system, the smartphone and IMU establish a session key with Elliptic Curve Diffie-Hellman (ECDH) over FourQ, then encrypt all communications with SNOW 3G (Santos-González et al., 2022). The Android implementation reported 417 ms for FourQ ECDH, compared with 721 ms for Curve25519 and 1876 ms for NIST P-256, and the design goal was specifically to avoid the traceability of the IMU devices when users are using them (Santos-González et al., 2022). The vision-assisted Wi‑Fi calibration framework makes a complementary privacy claim: the camera is used only during the data collection phase, and the visual pipeline is explicitly discarded after calibration so that deployment later requires only RSS sensing on end-user devices (Bilge et al., 26 Sep 2025).

The application space extends beyond navigation. Ubiquitous IPS work highlights worldwide seamless direction finding, first responders’ safety, and location-aware social networking (Youssef et al., 2012). RAVEL is explicitly motivated by expert museum guides and industrial settings that require sub-meter information (Papaioannou et al., 2023). A forensic computer-vision pipeline, Plug to Place, treats electrical sockets as indoor geographic markers standardized by country or region. Its three-stage pipeline uses YOLOv11 for socket detection, Xception for socket-type classification into 12 socket classes, and a rule-based socket-to-country mapping (Aftab et al., 18 Dec 2025). Reported performance includes mAP@0.5 = 0.843 for detection, accuracy = 0.912 for socket classification, and 96.29% country-level accuracy at >90% confidence on the TraffickCam subset of Hotels-50K (Aftab et al., 18 Dec 2025). The paper is explicit, however, that many images do not contain visible sockets and that high confidence thresholds improve accuracy but sharply reduce coverage (Aftab et al., 18 Dec 2025).

A common misconception is that indoor multimedia geolocation is synonymous with generic Wi‑Fi fingerprinting. The literature instead shows a heterogeneous field with room-level semantic labeling, panoramic visual retrieval, ground-texture localization, echoic SLAM, camera-radio multi-hypothesis tracking, mmWave layout sensing, and calibration pipelines that use vision only briefly. Another misconception is that sub-meter accuracy is a generic property of indoor systems. The reported results are modality- and setup-dependent: FM performs strongly in small confined areas but degrades at floor scale; camera systems may assume calibrated viewpoints; radio-map methods must handle noisy and sparse crowdsourced labels; and some pipelines depend on offline environment modeling or temporary calibration infrastructure (Popleteev et al., 2013, Papaioannou et al., 2023, Yi et al., 2023, Hu, 2020). This suggests that indoor multimedia geolocation is best understood as a family of coupled mapping, sensing, retrieval, and fusion problems whose outputs may be geometric, semantic, or forensic depending on the modality stack and deployment objective.