TL;DR
This paper introduces HPIPS, a smartphone-based system for high-precision indoor pedestrian positioning. It tackles the challenges of signal interference and data instability by fusing Wi-Fi Round-Trip Time (RTT) ranging with deep learning and an adaptive particle filter, achieving sub-meter accuracy.
What Problem Does This Paper Address?
Indoor positioning is crucial for many applications, but it's notoriously difficult. Unlike outdoor GPS, indoor environments suffer from signal blockage (non-line-of-sight) and reflections (multipath), making traditional radio-based methods unreliable. Existing fingerprinting techniques, which map signal strengths to locations, are sensitive to environmental changes, leading to fluctuating and discontinuous position estimates. This paper aims to overcome these limitations by developing a robust and accurate indoor positioning system that leverages readily available smartphone sensors and Wi-Fi technology.
Methods
The HPIPS system employs a two-pronged approach:
- Deep Convolutional Neural Network (CNN) for RTT Fingerprinting: To address the non-line-of-sight and multipath issues inherent in Wi-Fi signals, the authors propose using a CNN. Instead of traditional RSSI (Received Signal Strength Indicator) fingerprinting, they utilize Wi-Fi RTT ranging data. A novel approach is used to construct training datasets: fingerprint grayscale images are generated, incorporating specific Access Point (AP) locations. This allows the CNN to learn the complex, non-linear mapping between spatial positions and RTT ranging information, extracting representative spatial features for more accurate positioning.
- Adaptive Particle Filter for Sensor Fusion: To combat the large fluctuations and poor continuity of fingerprint-based positioning, an adaptive particle filter is introduced. This filter integrates data from the smartphone's MEMS sensors (accelerometer, gyroscope) and structured indoor map information. The MEMS data helps estimate pedestrian motion (step length, direction), while the map information acts as a constraint, preventing unrealistic "through-the-wall" predictions. This fusion improves the system's freedom and overall positioning accuracy and stability.
Datasets
The paper describes the construction of a custom dataset for training and evaluation. This dataset consists of Wi-Fi RTT ranging information collected in an 800 m² test environment. The data is visualized as grayscale images, incorporating AP locations and ranging data. No pre-existing public datasets are explicitly mentioned as being used.
Key Insights and Ideas
- RTT over RSSI for Fingerprinting: Wi-Fi RTT ranging is significantly more stable and less prone to environmental fluctuations than RSSI, making it a better candidate for fingerprint-based indoor positioning. The coefficient of variation (CV) analysis clearly demonstrates this.
- Image-based CNN for Spatial Features: Representing RTT data as grayscale images for CNN input allows the network to learn spatial relationships and extract richer features compared to traditional one-dimensional data processing.
- Adaptive Particle Filter for Robustness: Fusing MEMS sensor data and map constraints within an adaptive particle filter is crucial for smoothing out noisy CNN predictions and ensuring continuous, stable positioning, especially in challenging indoor environments.
- Map Constraints are Key: Incorporating building structure information (corridors, walls, doors) into the particle filter significantly improves accuracy by preventing unrealistic particle movements.
Key Results
The HPIPS system demonstrated impressive performance in an 800 m² test environment:
- High Accuracy: 94.2% of positioning errors were better than 1 meter.
- Low Average Error: The average positioning error was a mere 0.41 meters.
- Comparison: The system outperformed traditional CNN fingerprinting (average error 2.58m), PF+CNN fusion (average error 1.21m), and other common machine learning models (QDA, SVM, KNN) in terms of positioning error.
- Robustness: The system showed stability and continuity in both line-of-sight and non-line-of-sight scenarios, including through concrete walls and glass.
Limitations and Future Work
The authors acknowledge several limitations:
- Smartphone Anisotropy: The non-isotropic nature of smartphone antennas can lead to increased positioning errors if the phone's orientation changes. Users are currently required to maintain a fixed posture.
- Device Adaptability: The system's performance can be affected by variations in Wi-Fi signal reception capabilities across different commercial smartphones.
- Future Work:
- Investigating methods to mitigate the impact of smartphone attitude changes on positioning accuracy.
- Developing techniques to improve the algorithm's adaptability to diverse smartphone hardware.
- Exploring the use of deep learning to learn natural landmarks within indoor environments for further correction and refinement.
- Further leveraging building structure information and signal propagation models for enhanced accuracy.
Summary updated