TL;DR
This paper presents a novel indoor positioning system that combines smartphone MEMS sensor data with environmental "turning point" features detected by a deep learning model. This approach effectively calibrates the cumulative errors inherent in traditional pedestrian dead reckoning (PDR) methods, achieving significantly improved accuracy without requiring external infrastructure.
What Problem Does This Paper Address?
Accurate indoor positioning remains a significant challenge. While Global Navigation Satellite Systems (GNSS) are reliable outdoors, they are severely hampered by signal occlusion indoors. Existing indoor positioning methods often rely on external infrastructure (like Wi-Fi or Bluetooth beacons) or suffer from cumulative errors over time, especially with low-cost smartphone sensors. Pedestrian Dead Reckoning (PDR) is a promising sensor-based approach, but its accuracy degrades due to sensor noise and drift, making it unsuitable for long-term reliable positioning. This paper aims to bridge this gap by developing a robust, infrastructure-free indoor positioning system that leverages smartphone sensors and learned environmental cues.
Methods
The proposed system builds upon Pedestrian Dead Reckoning (PDR) by enhancing its accuracy and mitigating cumulative errors. The core components are:
- Improved PDR Modules: The paper details enhancements to the standard PDR pipeline, focusing on more accurate step detection, stride length estimation, and heading determination using smartphone MEMS sensors (accelerometer, gyroscope, magnetometer).
- Turning Point Detection using Deep Learning: A novel approach to identify "turning points" (e.g., corners, doorways) in indoor environments is introduced. This involves training a Stacked Denoising Auto-Encoder (SDAE) model on sensor data collected from various turning scenarios. The SDAE learns to extract robust features from noisy sensor data, enabling accurate classification of turning point categories.
- Particle Filter Fusion: A particle filter algorithm is employed to fuse the detected turning point information with the PDR positioning results. This fusion process acts as a calibration mechanism, correcting the accumulated drift in the PDR trajectory by leveraging the known locations of turning points as reference points.
Datasets
The paper describes the creation of custom datasets for training and evaluating the turning point detection model.
- Turning Point Dataset: Sensor data (orientation, gyroscope, geomagnetic field) was collected from 10 testers holding smartphones while walking through various turning areas in an indoor environment. This data was labeled into 8 direction categories.
- Evaluation Datasets: Experiments were conducted in two distinct indoor environments: a small-scale office (approx. 124 m²) and a large-scale conference hall (approx. 25 x 23 m²). These environments featured predefined walking paths with multiple turning points.
Key Insights and Ideas
- Environmental Cues as Calibration: The core innovation is using learned environmental features (turning points) as anchors for PDR calibration. This moves beyond relying solely on sensor data or external infrastructure.
- Deep Learning for Robust Feature Extraction: The use of SDAE for turning point detection demonstrates its effectiveness in extracting meaningful features from noisy smartphone sensor data, outperforming traditional machine learning classifiers in this context.
- Lightweight and Infrastructure-Free: The system's reliance on smartphone sensors and learned environmental features makes it inherently lightweight and eliminates the need for deploying additional hardware, making it highly scalable and cost-effective.
- Adaptive Calibration: The particle filter's ability to fuse PDR output with detected turning points allows for dynamic correction of accumulated errors, leading to sustained accuracy.
Key Results
- Turning Point Detection Accuracy: The proposed SDAE model achieved a high test accuracy of 98.2% for turning point detection.
- Small-Scale Environment: In a small office environment, the proposed PDR-calibrated system achieved an average positioning accuracy of 0.14 m, with 100% of errors less than 1 m. This is a significant improvement over the uncalibrated PDR (average error 0.83 m).
- Large-Scale Environment: In a large conference hall, the proposed system achieved an average positioning accuracy of 1.29 m, with 65% of errors less than 1.50 m. This is comparable to a commercial BLE calibration system but without the need for BLE beacon deployment.
- Smartphone Adaptability: The system demonstrated good adaptability across different smartphone models, with positioning errors generally less than 1 m for 70-97% of the test points, depending on the device.
Limitations and Future Work
- Dependence on Turning Points: The accuracy of the system is dependent on the density and detectability of turning points in the environment. In environments with few turning points, accuracy may degrade.
- Computational Efficiency: While designed to be lightweight, using a very large number of particles in the particle filter can reduce computational efficiency on smartphones.
- Future Work:
- Explore more nonlinear filtering methods (e.g., Extended Kalman Filter, Unscented Kalman Filter) to improve position estimation speed.
- Investigate new natural features (e.g., lighting, geomagnetic field variations, signals of opportunity) for calibration to achieve even higher accuracy and robustness.
- Consider more complex pedestrian activities and variations in smartphone holding poses for improved system adaptability.
Summary updated