IMU-Based Indoor Localization — Particle Filter & Bayesian Grid Filter

A Pedestrian Dead Reckoning (PDR) system for GPS-denied indoor environments, built on a Raspberry Pi 4 with a Sense HAT IMU. The system fuses gyroscope-based heading estimation with a 2D floor plan to track a user's position in real time using two probabilistic approaches: a Particle Filter and a Bayesian Grid Filter.

Developed as a group project for the Data Fusion Architecture (DFA) course, MSc Smart Systems Engineering, Hanze University of Applied Sciences (January 2026).

---

Team

Developed collaboratively by a 3-person team as part of the MSc Smart Systems Engineering programme at Hanze University.

My contributions focused on the Particle Filter implementation and the DSMS monitoring stack.

---

System Overview

Raspberry Pi (Sense HAT)
        │
        ├── imu_logger.py         # Raw IMU + joystick data logger
        ├── program2.py           # Live IMU publisher via MQTT
        │
        └── MQTT Broker (localhost)
                │
                ├── online_particle_pdr.py    # Particle Filter navigator
                ├── online_bayesian_pdr.py    # Bayesian Grid navigator
                │
                └── DSMS Stack
                        ├── program1.py         # CPU load publisher
                        ├── program3_fast.py    # Fast sliding window monitor
                        ├── program3_slow.py    # Slow sliding window monitor
                        └── program4.py         # Bernoulli sampler

---

Key Features

• Gyroscope bias calibration: 5-second stationary window at startup removes Zero-Rate Output (ZRO) drift
• Joystick-triggered step detection: Eliminates automated step detection errors; user clicks to register each step
• Two localization methods: Running on the same hardware for direct comparison
• Map-constrained filtering: Floor plan used as binary occupancy grid to reject physically impossible trajectories
• Kidnapped robot recovery: Particle Filter automatically resets if all particles hit walls simultaneously
• Real-time MQTT communication: Decoupled publisher/subscriber architecture between Pi and laptop
• DSMS monitoring stack: Lightweight system health monitor running in parallel with navigation

---

File Descriptions

Navigation & Sensing

File	Role
imu_logger.py	Runs on the Raspberry Pi. Reads gyroscope, accelerometer, and magnetometer data from the Sense HAT at ~100Hz and logs it to a CSV. Joystick direction keys register a step event; middle button stops logging. Used for offline data collection.
program2.py	Runs on the Raspberry Pi. Publishes live IMU data (orientation, accelerometer, gyroscope) to the MQTT broker on the sensors/sensehat topic at ~100Hz. This is the real-time data source for the online navigation scripts.
online_particle_pdr.py	Runs on the Raspberry Pi. Subscribes to the MQTT IMU stream and implements a Sequential Monte Carlo (Particle Filter) for real-time localization. Maintains 500 weighted particles, each with its own (x, y, θ) state. Applies a stop-on-wall collision model, multi-factor weight updates (map validity, stride consistency, heading consistency), and systematic resampling when effective sample size drops below N/2. Publishes position estimates back to MQTT and logs them to data/particle_run_log.csv.
online_bayesian_pdr.py	Runs on the Raspberry Pi. Subscribes to the MQTT IMU stream and implements a Discrete Grid-Based Bayesian Filter for real-time localization. Maintains a full probability distribution over a 2D grid of the floor plan. Applies prediction via grid shifting and Gaussian diffusion, and correction via map-based likelihood masking. Publishes position estimates to MQTT and logs them to data/bayesian_run_log.csv.
config.json	Central configuration file. Defines MQTT broker address, topics, map settings (filename, real-world width, cell size), PDR settings (step length, start position and heading), and particle filter parameters. Both online scripts read from this file at startup.

DSMS Monitoring Stack

File	Role
program1.py	Runs on the Raspberry Pi. Publishes CPU load (%) and temperature (°C) to the sensors/cpu MQTT topic every 10ms using psutil. Acts as the synthetic load generator for the DSMS pipeline.
program3_fast.py	Runs on the Raspberry Pi. Subscribes to sensors/cpu and computes a sliding window average with window size 5. Provides high-resolution, low-latency feedback sensitive to short spikes. Includes rule-based alerts for overload (>90%) and overheating (>75°C).
program3_slow.py	Same as above but with window size 50. Acts as a low-pass filter, smoothing out momentary spikes to reveal long-term CPU load trends.
program4.py	Runs on the Raspberry Pi. Subscribes to sensors/cpu and applies Bernoulli sampling, keeping only ~33% of incoming packets. Demonstrates that statistical properties (mean CPU load) can still be accurately estimated with significantly reduced data transmission — validating the DSMS bandwidth optimization hypothesis.

---

Results

Method	Map Consistency	Segment Wall Crossings	Endpoint Corridor Score
Bayesian Grid Filter	95.2%	1	116 px (mid-corridor)
Particle Filter	90.5%	2	35 px (near wall)

• Both filters ran in real time on Raspberry Pi 4 at ~18% total CPU load
• Particle Filter: computationally efficient, sensitive to heading drift
• Bayesian Grid Filter: stronger global map consistency, higher computational cost

---

Project Structure

imu-indoor-localization/
│
├── DFA_report-final.ipynb      # Full analysis notebook
├── DFA_report-final.pdf        # Academic report
├── map.png                     # Floor plan (binary occupancy map)
├── config.json                 # System configuration
│
├── online_particle_pdr.py      # Particle Filter (real-time)
├── online_bayesian_pdr.py      # Bayesian Grid Filter (real-time)
├── imu_logger.py               # IMU data logger (runs on Pi)
├── program1.py                 # DSMS: CPU load publisher
├── program2.py                 # DSMS: IMU data publisher
├── program3_fast.py            # DSMS: Fast sliding window monitor
├── program3_slow.py            # DSMS: Slow sliding window monitor
├── program4.py                 # DSMS: Bernoulli sampler
│
├── data/
│   ├── particle_run_log.csv
│   ├── bayesian_run_log.csv
│   ├── particle_session_fast.log
│   ├── particle_session_slow.log
│   ├── particle_session_sampled.log
│   ├── bayes_session_fast.log
│   ├── bayes_session_slow.log
│   └── bayes_session_sampled.log
│
├── requirements.txt
├── .gitignore
└── README.md

---

Installation

git clone https://github.com/MostafaJahanian/imu-indoor-localization.git
cd imu-indoor-localization
pip install -r requirements.txt

On Raspberry Pi additionally:

pip install sense-hat

---

Usage

Step 1 — Start the MQTT broker

mosquitto

Step 2 — On the Raspberry Pi

# Publish live IMU data
python program2.py

# Publish CPU monitoring data (separate terminal)
python program1.py

# Run DSMS monitors (separate terminals)
python program3_fast.py
python program3_slow.py
python program4.py

# Run navigation (separate terminal)
python online_particle_pdr.py
# OR
python online_bayesian_pdr.py

Step 3 — On the laptop

Open DFA_report-final.ipynb and run all cells. Logged data in data/ is used to reproduce all visualizations from the report.

---

Tech Stack

Python NumPy SciPy Pandas Matplotlib Paho-MQTT Sense HAT Raspberry Pi Particle Filter Bayesian Filter MQTT asyncio

---

Academic Report

Full technical report covering system architecture, mathematical models, experimental results, sensitivity analysis, and hardware performance is available in DFA_report-final.pdf.

---

Author

Mostafa Jahanian LinkedIn · GitHub