MINOR PROJECT PRESENTATION
Wi-Fi Sensing
Algorithm
Development for
Indoor
Localization and
Human Activity
Recognition
TABLES OF CONTENT
• What is Indoor Localization?
• Why Indoor Localization?
• Machine Learning Algorithms for Indoor Localization for Predication of
Location of Users and Steps Taken by the Users.
• Model Fusion for Indoor Localization
• Wi-Fi and BLE Data Fusion for Indoor Localization
• Conclusion
• Future Scope
WHAT IS INDOOR
LOCALIZATION?
•Definition: Indoor localization refers to determining the location of an
object or person within a building or enclosed environment.
•Objective: To provide precise coordinates or positional data such as
building, floor, room, or relative position.
•Technology: Utilizes technologies like Wi-Fi, Bluetooth, RFID, UWB,
visible light communication (VLC), inertial sensors, or camera-based
systems.
•Output: Provides position in terms of: - Coordinates (latitude,
longitude, height/floor level) - Contextual information (e.g., room or
zone within a building)
•Practical Need- WHY INDOOR LOCALIZATION?
•GPS Limitations: GPS signals are weak or unavailable inside buildings due to signal attenuation.
•High-precision Requirements: Indoor environments often need sub-meter accuracy for
navigation and tracking.
•Applications-
•Navigation and Guidance: - Assist users in malls, airports, hospitals, and museums with turn-by-
turn directions.
•Help visually impaired individuals navigate indoor spaces.
•Asset Tracking: - Monitor inventory, equipment, or personnel in warehouses and factories. Track
medical devices and patients in hospitals.
•Emergency Services: - Locate individuals during emergencies like fires or earthquakes.
•Smart Buildings :- Enable automated control of HVAC, lighting, and security based on occupant
location.
•Retail and Marketing : - Deliver location-based advertisements and promotions in malls or
stores.
MACHINE LEARNING ALGORITHMS FOR INDOOR
LO C A L I Z AT I O N F O R P R E D I C AT I O N O F LO C AT I O N
O F U S E R S A N D S T E P S TA K E N BY T H E U S E R S .
K Nearest Neighbors
1. Data Collection:
• RSSI fingerprints are collected at predefined reference points (RPs) across the indoor environment. Each fingerprint is
associated with a known location and represents the RSSI values from visible APs.
2. Feature Representation:
• The RSSI fingerprints are represented as feature vectors, where each dimension corresponds to an AP's RSSI value.
Missing RSSI values (e.g., when an AP is out of range) are typically replaced by a constant, such as -100 dBm, to maintain
uniformity.
3. Distance Metric:
• The algorithm computes the similarity between the target RSSI fingerprint and all stored fingerprints using a distance
metric.
• Euclidean Distance: Effective for continuous features but sensitive to scaling.
4. Neighbor Selection:
• The k nearest neighbors is identified based on the chosen distance metric. The value of k (Here k=5) significantly impacts
performance: small k values may lead to overfitting, while large k values can smooth out class boundaries excessively.
5. Prediction:
• The algorithm predicts the location of the target fingerprint based on the majority class (classification) or the average
location (regression) of the k neighbors.
Bayesian Algorithm for Indoor Localization-
Principle: Based on Bayes' theorem, it computes the posterior probability of a location given observed data.
Input: RSSI, time-of-flight (ToF), or other sensor data as evidence.
Output: Probabilistic estimation of the user's position.
Advantages: - Handles uncertainty in noisy environments effectively.
Suitable for probabilistic and dynamic systems: - Can incorporate prior knowledge about the environment (e.g.,
floor plans or signal maps).
Limitations: - Computationally intensive for large spaces.
Requires accurate prior probabilities and observation models.
Use Cases: Ideal for fusing data from multiple sensors and handling uncertainty in signal measurements.
Convolutional Neural Networks (CNN) for Indoor Localization-
Principle: Uses convolutional layers to extract spatial features from input data (e.g., RSSI maps, images, or
fingerprinting data).
Input: Signal strength maps, Wi-Fi RSSI data, or images of the environment.
Output: Predicted location coordinates or room labels.
Advantages: - Captures spatial patterns in RSSI or signal maps. High accuracy with sufficient training data. Can
generalize well for complex environments.
Limitations: - Requires a large dataset for training. High computational cost for training and inference.
Use Cases: Used in systems where spatial signal patterns are key, such as fingerprinting-based localization.
Deep Neural Networks (DNN) for Indoor Localization-
Principle: Multi-layer perceptrons (MLPs) or other deep architectures map high-dimensional input data
to location outputs. Input: RSSI values, sensor fusion data, or other features.
Output: Coordinates or classification of zones/rooms.
Advantages :- Handles complex, nonlinear relationships between input signals and locations.
Adaptable to diverse data types (e.g., Wi-Fi, BLE, or UWB). Can combine different data sources (e.g.,
RSSI + IMU data).
Limitations :- Needs significant computational resources. Risk of overfitting with limited data.
Use Cases :- Suitable for environments with diverse and complex signal dynamics, supporting both
regression (coordinates) and classification tasks.
Long Short-Term Memory (LSTM)
Type: A variant of Recurrent Neural Networks (RNNs).
Purpose: Designed to capture and model long-term dependencies in sequential data.
Key Feature: Addresses the vanishing gradient problem in standard RNNs by maintaining a memory
cell.
Why LSTM for Indoor Localization?
Recurrent Neural Network for Indoor Localization (RNN)
•Sequential Data: RNNs handle time-series data like RSSI effectively for indoor localization.
•Temporal Awareness: They capture dependencies between past and current signal patterns.
•Sensor Fusion: RNNs combine data from Wi-Fi, Bluetooth, and sensors for improved accuracy.
•Dynamic Environments: They adapt well to changing indoor layouts and signal conditions.
•Enhanced Models: LSTMs and GRUs address RNN limitations for better performance.
Random Forest for Indoor Localization
• Feature Importance: Random Forest identifies key features like RSSI and floor ID for localization.
• Robust Performance: It handles noisy and missing data common in indoor environments.
• Multi-Source Data: Combines data from Wi-Fi, Bluetooth, and IMU sensors efficiently.
• Parallel Processing: Random Forest trains on multiple trees, ensuring faster and accurate
predictions.
RESULTS OF INDOOR LOCALIZATION USING
KNN, BAYESIAN, CNN AND DNN ALGORITHM
Algorithms KNN Bayesian DNN CNN
Parameters
Accuracy (in %) 97.456 99.348 96.313 43.34
Precision (in %) 97.58 99.441 96.347 43.24
F1 Score / R2 Score 97.15 99.37 94.505 46.09
Recall 96.843 99.318 96.313 50.5
Hamming Loss 0.048 0.652 0.03686 56.66
MODEL FUSION RESULTS
Parameters KNN + LTSM KNN + BAYESIAN
Accuracy(in %) 63.70036 85.30096
Precision(in %) 64.29873 86.529873
Recall 0.637003 0.8503090
F1 Score 0.629914 0.8499749
Hamming Loss 0.362996 0.1496990
Mean Squared Error 0.313015 0.2718154
Mean Absolute Error 0.223520 0.1835506
Computation Time(s) 377.26 1.4211
RESULTS OF WI-FI RSSI AND BLE RSSI DATA FUSION
USING KNN, RNN AND RANDOM FOREST MODELS
Models KNN RNN Random Forest
Regression
Parameters
Mean Squared Error (MSE) 36.2805 64.957 1.585
Mean Absolute Error (MAE) 1.1697 4.838 0.28055
Accuracy (in %) 63.7194 35.042 98.4148
This research for indoor localization using Data Fusion Model is compatible and universal for various kind of
technologies and multiple data sources can be integrated to produce more consistent, accurate, and useful
information than that provided by any individual data source.
USER PLOT: LATITUDE VS
LONGITUDE FOR USER POSITION
CONCLUSION
Importance- Crucial for applications where GPS is ineffective, such as malls, hospitals,
and airports.
Enhances navigation, asset tracking, and emergency response capabilities.
Technological Advancements: - Integration of wireless technologies (Wi-Fi, BLE, UWB) and
advanced algorithms (ML/DL models) has improved accuracy.
Challenges : - Signal interference and environmental dynamics (e.g., furniture, walls)
remain key hurdles.
Privacy and security concerns when tracking individuals indoors.
Impact : - Revolutionizing industries like retail, healthcare, logistics, and smart buildings.
Boosts user convenience and operational efficiency.
FUTURE SCOPE
Technological Advancements:
AI and ML Integration: - Use of LSTMs, transformers, and hybrid models for better
prediction and context understanding. - Enhanced model fusion combining signals from
diverse sources (WiFi, BLE, RFID, sensors).
Sensor Fusion: - Combining data from inertial sensors, cameras, and wireless signals for
higher precision.
Edge Computing: - Real-time localization through on-device processing for reduced latency
and privacy protection.
Ultra-Wideband (UWB): - Promises centimeter-level accuracy with minimal interference.
Applications Expansion:
AR/VR: - Immersive gaming and navigation experiences driven by precise indoor tracking.
Autonomous Systems: - Drones and robots for delivery, cleaning, and monitoring in indoor
environments.
Smart Cities: - Seamless integration with IoT for personalized services in urban spaces.
Healthcare: - Advanced patient tracking, emergency assistance, and monitoring in
hospitals.
T H A N K YO U
PR E S E N T E D BY:
KU M A R I P R IYA S H A R M A
(1 0 0 1 3 2 0 2 8 2 1 )
A D IT YA G U P TA (0 8 7 1 3 2 0 2 8 2 1 )
