AmI · Lecture 5 — Indoor Localization

Indoor Positioning Systems (IPS)

Definition

Indoor Localization is the process of obtaining the location of a user or a device in an indoor environment. Systems based on it are Indoor Positioning Systems (IPS).

Made possible by the proliferation of mobile/wearable devices with wireless communication capabilities. Indoor location is an important context information for Ambient Intelligence systems (links back to context-awareness in Lecture 1).

The problem

Given a map of the building (walking areas + obstacles) and a reference coordinate system, the goal is to determine the coordinates (x, y) where an entity (person, object, robot) is located at a given instant. It is usually framed as a regression problem.

Applications

🧭

Navigation

Continuously detect the user's location and compute the shortest path to a destination in smart buildings.

🏠

Smart Homes (AAL)

Ambient Assisted Living: monitor elderly subjects' location & indoor trajectories. Hard in multi-inhabitant settings.

🔑

Object Localization

Commercial solutions (e.g., AirTag) locate objects.

Why not simply use GPS?

GPS is great outdoors but not suitable indoors: signal strength is low, and indoor structures (roof, walls, objects) attenuate and scatter it.
Average GPS error: 5–10 meters — far too coarse indoors (ideally we want < 1 m).

Indoor vs Outdoor

Reduced coverage area (a building)
Less impact of weather
Resources available (electricity, Internet)
Slower movement (walking speed)
Buildings usually private (planimetry not public)

Active vs Passive

Active: the device computes its position while moving, possibly relying on sensors in the environment (the focus of this class)
Passive: a distributed system computes the position without explicit cooperation of the subject (e.g., cameras)

Characteristics of an IPS

An IPS is described along many dimensions: accuracy, coverage, cost, update rate, latency, privacy, robustness, scalability, infrastructure…

Accuracy

Degree of conformance of an estimated position to the true value. Two common metrics:

RMSD = √( (1/n) Σ (P̂ᵢ − Pᵢ)² ) · AAPD = (1/n) Σ |P̂ᵢ − Pᵢ|

Root Mean Squared Deviation (RMSD) — good if the data distribution is normal.
Average Absolute Position Deviation — more robust to outliers.

Coverage

Local: small, well-defined, not extendable area.
Scalable: coverage can grow with additional hardware.
Global: worldwide (only applicable to GPS).

Update Rate

Periodic — regular updates (e.g., Hz)
On request — triggered by user/remote device
On event — when a specific event occurs

System Latency

Real-time — delay not perceived (crucial for navigation)
Best effort — answer as soon as possible
Post processing — analysis done after collection

Data output may differ by application: position, heading/bearing, acceleration, speed/velocity. For IPS there is no single-technology solution — a high diversity of solutions exists, all requiring dedicated local infrastructure.

Radio Technologies for IPS

Most proposed solutions are based on radio communication technologies (WiFi, Bluetooth, RFID, Zigbee). They compute the position of a user/device, and require one or more antennas in the environment.

Major challenges

Signal reflection on walls/furniture, density of obstacles, dynamic environments (people moving, doors opening) — all degrade accuracy. Ideally positioning error should be below 1 meter.

Technology	Key points
WiFi (802.11)	Widely adopted, no ad-hoc infrastructure needed (APs everywhere) — but APs are deployed for communication, not localization, so ad-hoc algorithms are needed to improve accuracy.
Bluetooth Low Energy (BLE)	More accurate and lower energy than WiFi; uses standard protocols (iBeacons, Eddystone).
BLE Beacons	Small transmitters continuously broadcasting a UUID to nearby devices. Conceived for proximity, now used for localization. More accurate than WiFi, but deployment/maintenance harder (how many, where, what height, battery replacement).
Ultra WideBand (UWB)	Short pulses over a large bandwidth, low duty cycle. Less affected by interference, can penetrate obstacles, less sensitive to multipath. Drawbacks: costly/complex infrastructure, not yet standard in personal devices. Uses anchors (fixed, known positions) + a mobile tag; more accurate than BLE but costlier.

⭐ RSSI-based Localization (Simulation Q6)

This is the exact subject of simulation Q6. Nail what RSSI is, what affects it, and what a single value can vs cannot tell you.

Received Signal Strength Indicator (RSSI)

RSSI is the signal power (strength) received from a radio antenna — usually considered for WiFi and BLE (not UWB). It can estimate the distance d of a receiver from an antenna, but estimation accuracy drops as distance increases, and it is heavily affected by interferences.

d ≈ 10 ^ ((measuredPower − RSSI) / (10·N))

measuredPower = RSSI at 1 meter · N = environmental factor

RSSI is a time series (in decibels)

RSSI data is expressed in decibels (dB) → negative values (small positive milliWatt energy on a logarithmic scale).
The closer to 0, the stronger the signal.

Trilateration / Multilateration

Each RSSI value → a distance → defines a radius (circle) around an antenna.
With three RSSI values you can intersect the circles to locate the subject → trilateration.
With more than three values it's called multilateration.

Problem of RSSI distance-based approaches

Computing distance from RSSI is not accurate — independently of the technology — because of signal interferences in the indoor environment. Hence more sophisticated approaches (fingerprinting, CSI…) have been proposed.

ThinkSimulation Q6 asks which RSSI statement is FALSE. Why is "RSSI values can only be obtained using Bluetooth Low Energy beacons" the false one?

Answer: Because RSSI is the received signal power from any radio antenna — the slides state it is "usually considered for WiFi and BLE." So it is not exclusive to BLE beacons. The other statements are true: RSSI is subject to the multi-path effect, a single RSSI value can only estimate distance (not full position), and distance-based localization needs at least 3 values (trilateration).

Fingerprinting

The rationale

For each specific point in the environment, RSSI signals are characteristic — it's unlikely the same set of RSSI values from different antennas repeats in different locations. A fingerprint at a position is a set of tuples <ID, RSS_ID> (antenna ID + the value it produces there; you can aggregate several readings, e.g. average).

① Offline phase (training)

A person/robot moves through the building
At each spot: report the true position + collect a fingerprint
Repeated across the environment (e.g., each meter)
Result stored server-side = radio map: a set of <Position, Fingerprint> pairs

② Online phase (inference)

The user is in the environment; position is unknown
Collect current RSSI values
Basic approach: pick the fingerprint in the radio map closest to current RSSI
Its position = the inferred position

"Naive" approach — minimize Euclidean distance

x̂ = argmin_xⱼ Σᵢ ( rᵢ − ρᵢ(xⱼ) )²

rᵢ = detected RSSI signal · ρᵢ(xⱼ) = fingerprint value · x̂ = inferred position

Fingerprinting based on Machine Learning

In the same position the fingerprint may change (noise, environmental changes, presence of people) → naive approaches fail in noisy real-world scenarios.
ML models are preferred for robustness, usually framed as a regression problem.
Models: K-Nearest Neighbor (common baseline — match the K closest fingerprints and average their positions), classic ML (SVM, Random Forest, Bayesian), and Deep Neural Networks.
CNN approach: convert the RSSI values at a time instant into an image (pixel intensity = signal strength per antenna at a location) → feed a CNN. (Ties back to CNNs in Lecture 3.)
LSTM + self-attention + CNN: each input is a sequence of fingerprints; LSTM+attention map sequences into embeddings, CNN captures spatial properties. (Combines Lectures 3 & 4 — LSTM, self-attention, CNN — and the lecture notes this combination is very common and effective.)

ThinkWhy is fingerprinting often preferred over RSSI distance-based trilateration, and what is its main practical drawback?

Answer: Trilateration relies on converting RSSI → distance, which is inaccurate due to indoor interference/multipath. Fingerprinting instead matches the whole RSSI pattern to a pre-built radio map, sidestepping the distance estimate, and ML models add robustness to noise. The drawback: it requires a costly offline survey, and new fingerprints are needed whenever the space changes even slightly.

Channel State Information (CSI)

CSI — an alternative to RSSI

RSSI is widely used due to its simplicity, but it is susceptible to multipath effects and interference. CSI is obtained from WiFi access points and mitigates these problems — it's more robust, but only a few off-the-shelf receiver devices can actually compute it.

CSI encodes the phase and amplitude of the received signal at different frequencies, implicitly capturing the scattering/refraction the signal encountered.

OFDM (Orthogonal Frequency-Division Multiplexing)

The transmitter (WiFi AP) encodes each message on multiple subcarrier frequencies within one channel, reducing interference and optimizing bandwidth.
For each subcarrier we learn how the signal propagates (amplitude & phase). The receiver needs an OFDM demodulator to decompose the signal into its subcarriers.
Each location has unique patterns — distance & interference affect amplitude/phase of each frequency in a specific way.

RSSI

One aggregate strength value (LoS + all paths summed)
Simple, cheap, broadly available
Prone to multipath fading

CSI

Per-subcarrier amplitude & phase — separates the paths
More reliable information than RSSI
Not easily available on off-the-shelf NICs

CSI-based localization is usually done with supervised fingerprinting — similar techniques to those used for RSSI signals (OFDM demodulator → CSI extraction → CSI processing → fingerprint matching).

Other Localization Approaches

Angle & time based (advanced anchors, usually UWB)

Technique	Idea & trade-off
Angle of Arrival (AoA)	Anchors compute the angle to the receiver via directionally-sensitive antennas. Combining distances + angles → sub-meter accuracy. Drawback: sophisticated/costly hardware.
Time of Arrival / Flight (ToA/ToF)	Measure the absolute travel time of the signal (distance = travelTime × waveSpeed). Needs precise clock synchronization between antennas and receiver; works best in Line-of-Sight. Usually UWB.
Time Difference of Arrival (TDoA)	Improves ToA by removing the mobile device's clock bias — uses the time difference between antenna pairs + multilateration. Drawback: antennas must be perfectly synchronized with each other.

Quick pros / cons recap

RSSI: easy/cheap, many technologies — but prone to multipath, lower accuracy. CSI: robust to multipath — but not on off-the-shelf NICs. AoA: high accuracy, no fingerprinting — but complex hardware, degrades with distance. ToF: high accuracy — but needs clock sync + LoS. TDoA: no device-clock sync needed — but needs anchor sync. Fingerprinting: easy to use — but new fingerprints needed on any change.

Inertial approaches — Dead Reckoning

Inertial sensors (accelerometer + gyroscope) capture acceleration in a direction. Starting from a position obtained by another method (e.g., RSSI), estimate displacements — useful when positioning isn't constantly available.
Pipeline: step detection → step-length estimation (a personalized step model helps) + heading estimation → position update (analytically or via DL).
Problem: inertial sensors are noisy → drift grows linearly with time; small early errors compound. So dead reckoning is for small displacements and cannot be used alone.

Alternative (non-radio) IPS technologies

🚶

Passive InfraRed (PIR)

Detects human motion in an area, no wearable needed. Coarse-grained, one sensor per area, hard with multiple users (can count, not track individuals).

📷

Computer Vision

Mobile camera: needs markers, active framing → good for navigation. Static cameras: passive, easier, but intrusive (privacy) → good for surveillance.

💡

Visible Light Comm. (VLC)

Data emitted by LEDs; receiver's light sensor estimates position/direction. Pro: LED proliferation. Con: requires Line-of-Sight.

✓

Final Quiz — Exam Style

9 MCQs + 2 open-ended. Q1 is the simulation's Q6 reproduced verbatim; the rest are style-matched to the exam.

⭐ Simulation Q6. Which of these sentences on RSSI values is FALSE?

ARSSI values can only be obtained using Bluetooth Low Energy beacons

BRSSI values are subject to the multi-path effect

CIt is necessary to have at least 3 RSSI values for accurate localization using distance-based approaches

DA single RSSI value can only estimate the distance from the receiver and the transmitter antennas, but it does not provide sufficient information for localization

A is false — RSSI is the received power from any radio antenna, "usually considered for WiFi and BLE." It is not exclusive to BLE beacons. B, C, D are all true.

Why is GPS not suitable for indoor localization?

AGPS only works at night

BGPS has sub-centimeter accuracy that is too precise indoors

CIts signal is weak and indoor structures attenuate/scatter it, giving 5–10 m error

DGPS requires a dedicated BLE beacon in every room

C — GPS is for outdoors; walls/roof/objects attenuate and scatter the weak signal, yielding 5–10 m error (too coarse for the <1 m needed indoors).

In RSSI-based localization, what is trilateration?

AMeasuring the angle of the signal with respect to three antennas

BUsing three RSSI-derived distances as circle radii and computing their intersection to locate the subject

CSynchronizing three clocks to measure signal travel time

DConverting RSSI values into an image for a CNN

B — each RSSI → a distance → a circle radius around an antenna; intersecting three circles gives the position. With more than three values it's called multilateration.

What is a fingerprint in fingerprinting-based localization?

AThe biometric scan of the user's finger used to authenticate

BThe angle of arrival of the signal at a single antenna

CThe GPS coordinate stored for each room

DA set of tuples ⟨antenna ID, RSSI value⟩ characteristic of a specific position

D — each position has characteristic RSSI values from the various antennas; a fingerprint is the set of ⟨ID, RSS_ID⟩ tuples. The radio map stores ⟨Position, Fingerprint⟩ pairs.

What happens in the offline phase of fingerprinting?

AA person/robot moves through the building, reporting true positions and collecting fingerprints to build the radio map

BThe user's unknown position is inferred from current RSSI values

CThe BLE beacons' batteries are recharged

DThe GPS satellites are calibrated

A — the offline (training) phase builds the radio map of ⟨Position, Fingerprint⟩ pairs. The online phase then infers the unknown position by matching current RSSI to the map.

What is the main advantage of CSI over RSSI?

ACSI is available on every off-the-shelf device

BCSI requires no infrastructure at all

CCSI captures per-subcarrier amplitude and phase, so it is more robust to multipath and interference

DCSI is a single aggregate value, making it simpler than RSSI

C — CSI encodes amplitude and phase across subcarriers (via OFDM), separating signal paths and mitigating multipath. The trade-off: only a few off-the-shelf NICs can compute it.

Which statement about UWB is correct?

AUWB is more affected by multipath than WiFi and BLE

BUWB uses fixed anchors and a mobile tag, is less sensitive to multipath and can penetrate obstacles, but needs costly infrastructure

CUWB is the standard RSSI technology used in all smartphones

DUWB cannot achieve sub-meter accuracy

B — UWB sends short pulses over a large bandwidth: less interference, penetrates obstacles, less multipath, and can reach sub-meter accuracy (esp. with AoA/ToF), but the infrastructure is costly and not yet standard in personal devices.

Why does Dead Reckoning need to be combined with another localization technique?

AIts drift grows linearly with time, so small early errors compound — it can only estimate small displacements

BIt only works outdoors with GPS

CIt requires a fingerprint radio map to function at all

DIt cannot detect steps or heading

A — inertial sensors are noisy; the position error (drift) accumulates over time. So dead reckoning estimates short displacements from a position obtained by another method (e.g., RSSI) and cannot be used alone.

Which statement about Time Difference of Arrival (TDoA) is correct?

AIt measures the angle of the signal at each antenna

BIt requires the mobile device's clock to be synchronized with the antennas

CIt converts RSSI into an image for a CNN

DIt uses the time difference between antenna pairs to remove the device's clock bias, but requires the antennas to be synchronized with each other

D — TDoA improves ToA by using time differences between antenna pairs (removing the device clock bias) + multilateration, but it needs the anchors/antennas to be perfectly synchronized among themselves.

Open-ended · RSSI & Fingerprinting

a) What is RSSI, what affects its accuracy, and how can it be used (with trilateration) to localize a device?

b) Explain fingerprinting: the offline and online phases, and why ML models are often preferred over the naive approach.

Model answer

a) RSSI (Received Signal Strength Indicator) is the signal power received from a radio antenna, usually for WiFi and BLE, expressed in dB (negative values; closer to 0 = stronger). It can estimate the distance d between receiver and antenna (d ≈ 10^((measuredPower−RSSI)/(10·N))), but accuracy drops with distance and is heavily affected by interference / multipath. With three RSSI values, each distance defines a circle radius around an antenna; intersecting the three circles localizes the subject — this is trilateration (multilateration with more than three). Because the distance estimate is inaccurate, this is unreliable in real indoor environments.

b) Fingerprinting exploits that each position has characteristic RSSI values. Offline phase: a person/robot surveys the building, at each spot reporting the true position and collecting a fingerprint (set of ⟨antenna ID, RSSI⟩ tuples); the result is a radio map of ⟨Position, Fingerprint⟩ pairs. Online phase: the user's RSSI is collected and the closest fingerprint in the radio map is chosen — its position is the inferred one (naive approach = minimize Euclidean distance between detected RSSI and fingerprints). Because the fingerprint at a position changes due to noise/people/environmental changes, the naive approach fails in noisy scenarios; ML models (KNN, SVM, Random Forest, DNN/CNN) are preferred for robustness, framing localization as a regression problem.

Open-ended · Localization technologies comparison

a) Compare WiFi, BLE beacons and UWB as radio technologies for IPS (accuracy, energy, infrastructure).

b) Briefly describe AoA, ToF and TDoA and the main requirement/drawback of each.

Model answer

a) WiFi is widely available (APs everywhere, no ad-hoc infrastructure), but APs are made for communication not localization, so ad-hoc algorithms are needed and accuracy is limited. BLE beacons are more accurate and lower-energy than WiFi, broadcasting a UUID; but deployment/maintenance is harder (how many, where, what height, battery replacement). UWB sends short pulses over a large bandwidth: less interference, penetrates obstacles, less multipath, and most accurate (sub-meter); it uses fixed anchors + a mobile tag, but the infrastructure is costly/complex and not yet standard in personal devices.

b) AoA (Angle of Arrival): anchors measure the angle to the receiver via directional antennas; combining angles + distances gives sub-meter accuracy, but needs costly/sophisticated hardware and degrades with distance. ToF/ToA (Time of Flight/Arrival): measures the absolute signal travel time (distance = time × wave speed); high accuracy but requires precise clock synchronization between antennas and device, and Line-of-Sight. TDoA (Time Difference of Arrival): uses the time difference between antenna pairs to remove the device's clock bias, then multilateration; it removes the need for device-clock sync, but the antennas must be perfectly synchronized with each other.

0/9