An Adaptive Method for Switching between Pedestrian/Car Indoor Positioning Algorithms based on Multilayer Time Sequences

Zhining Gu, Wei Guo, Chaoyang Li, Xinyan Zhu, Tao GuoSensors 2018 · DOI

TL;DR

This paper introduces an adaptive system that intelligently switches between pedestrian and car indoor positioning algorithms based on detected user behavior. By leveraging multilayer time sequences (MTSs) to filter noisy sensor data and optimize state transitions, the system significantly improves localization accuracy and reduces switching delays compared to traditional methods.

What Problem Does This Paper Address?

Accurate indoor positioning is crucial for location-based services, but existing methods often struggle with diverse user behaviors. Pedestrian Dead Reckoning (PDR) works well for walking but fails for driving, while methods like Bluetooth positioning are better suited for vehicles but less efficient for pedestrians. Relying on a single algorithm is suboptimal, leading to inaccuracies or delays when user states change (e.g., from walking to driving). The challenge lies in seamlessly and accurately identifying these state changes in real-time to switch to the most appropriate positioning algorithm, especially in complex indoor environments where sensor data can be noisy.

Methods

The core of the proposed solution is an adaptive method based on Multilayer Time Sequences (MTSs). This approach combines two main strategies:

  • Deep Learning for State Recognition: A Convolutional Neural Network (CNN) is used to process sensor data (accelerometer, gyroscope) to initially identify user states (static, walking, in-car, other). This leverages the CNN's ability to extract features without requiring domain-specific knowledge.
  • Multilayer Time Sequence (MTS) Processing: This is the novel contribution. It involves two layers:
  • Smaller-scale Layer (Noise Filtering): This layer uses sliding time plots and a majority voting rule to filter out noisy data from the initial CNN output. It considers both a priori (previous) and posteriori (current) contexts to refine the state identification.
  • Larger-scale Layer (State Chain Optimization): This layer employs a "state chain" concept to further reduce noise and, crucially, minimize the time delay associated with switching between positioning algorithms. It models the influence of previous states on the current one and uses a priority strategy to resolve inconsistencies between CNN and Bluetooth-based velocity estimations.

The system ultimately combines the outputs from the CNN (behavior model) and Bluetooth positioning (velocity data) to determine the final state and trigger the appropriate positioning algorithm (PDR for walking, Bluetooth for in-car).

Datasets

The paper describes data collection using smartphones (HUAWEI mate9, XIAOMI NOTE, MEIZU) equipped with inertial sensors. Data was collected in a controlled experimental site (a parking lot at Wuhan University) with deployed Bluetooth devices. The collected data includes accelerometer and gyroscope readings, as well as Bluetooth positioning data. The authors do not mention using any pre-existing public datasets for training or evaluation; rather, they collected their own data for this study.

Key Insights and Ideas

  • Behavior Context is King: The most significant insight is that accurately recognizing the user's behavior is the key to selecting the right indoor positioning algorithm. Simply relying on raw sensor data or a single positioning technology is insufficient.
  • Two-Pronged Approach for Robustness: Combining deep learning for initial state detection with a multilayer sequence analysis for noise filtering and state transition optimization creates a more robust system than either approach alone.
  • Time Delay is a Critical Factor: The paper emphasizes that not only accuracy but also the speed at which the system switches algorithms is vital for a good user experience. The state chain optimization directly addresses this.
  • MTS as a Unified Framework: The MTS framework elegantly handles both fine-grained noise filtering and coarse-grained state transition optimization, providing a cohesive approach to adaptive positioning.

Key Results

The proposed MTS method demonstrated significant improvements over a baseline CNN method:

  • Accuracy Improvements:
    • Static: 5.5% average improvement
    • Walking: 45.47% average improvement
    • In-car: 26.23% average improvement
    • Other states: 21% average improvement
  • Time Delay Reduction:
    • Transition between states: Reduced by approximately 0.5–8.5 seconds.
  • Entire process (multistate events): Reduced by approximately 24 seconds.

Limitations and Future Work

  • Noisy Data Persistence: The authors acknowledge that while their method significantly reduces noisy data, it cannot be completely eliminated.
  • Time Delay Optimization: Similarly, the time delay cannot be entirely removed, though it is substantially reduced.
  • Future Work: The paper suggests that further research is needed to improve upon existing related methods as a basis for future advancements. The authors also mention that the "state chain" method could be further refined.

Summary updated

Abstract

Pedestrian dead reckoning (PDR) positioning algorithms can be used to obtain a target's location only for movement with step features and not for driving, for which the trilateral Bluetooth indoor positioning method can be used. In this study, to obtain the precise locations of different states (pedestrian/car) using the corresponding positioning algorithms, we propose an adaptive method for switching between the PDR and car indoor positioning algorithms based on multilayer time sequences (MTSs). MTSs, which consider the behavior context, comprise two main aspects: filtering of noisy data in small-scale time sequences and using a state chain to reduce the time delay of algorithm switching in large-scale time sequences. The proposed method can be expected to realize the recognition of stationary, walking, driving, or other states; switch to the correct indoor positioning algorithm; and improve the accuracy of localization compared to using a single positioning algorithm. Our experiments show that the recognition of static, walking, driving, and other states improves by 5.5%, 45.47%, 26.23%, and 21% on average, respectively, compared with convolutional neural network (CNN) method. The time delay decreases by approximately 0.5-8.5 s for the transition between states and by approximately 24 s for the entire process.
BibTeX
@inproceedings{gu2018adaptive,
  author = {Zhining Gu and Wei Guo and Chaoyang Li and Xinyan Zhu and Tao Guo},
  title = {{An Adaptive Method for Switching between Pedestrian/Car Indoor Positioning Algorithms based on Multilayer Time Sequences}},
  booktitle = {Sensors},
  year = {2018},
  doi = {10.3390/s18030711},
}
Figure