Smartphone Location-Independent Physical Activity Recognition Based on Transportation Natural Vibration Analysis

Tae Ho Hur, Jae Hun Bang, Dohyeong Kim, Oresti Banos, Sungyoung Lee2017 · DOI

TL;DR

This paper introduces a novel smartphone-based activity recognition system that leverages the natural vibrations of buses and subways to distinguish between these transportation modes, overcoming a common challenge in existing methods. By combining accelerometer and gyroscope data for orientation independence, natural vibration feature extraction, and a correction algorithm, the system achieves robust and energy-efficient activity recognition in real-world conditions.

What Problem Does This Paper Address?

Recognizing human activities using smartphones (Activity Recognition or AR) is a growing field with applications in healthcare, wellness, and life-logging. While accelerometers are a popular sensor for AR, they struggle to accurately differentiate between similar static activities like standing, riding a bus, or riding a subway. This is because these activities often involve minimal user movement, leading to similar acceleration patterns. Existing methods that rely on orientation-independent features often sacrifice the ability to distinguish between these nuanced transportation modes. This paper aims to bridge this gap by analyzing the unique natural vibrations of buses and subways.

Methods

The proposed system employs several key techniques:

  • Orientation and Position Independence: The system uses the gyroscope to align the accelerometer with the direction of gravity. This ensures that the sensor data is not affected by how the user holds or carries their smartphone, making it orientation and position-independent.
  • Natural Vibration Feature Extraction: To distinguish between buses and subways, the researchers analyze the natural vibrations inherent to these vehicles. They use a Fast Fourier Transform (FFT) on the Z-axis accelerometer signal (after gravity removal) to transform the time-domain data into the frequency domain, revealing distinct vibration frequencies for buses and subways. Additional features are extracted from the X and Y axes and their correlations with the Z-axis.
  • Feature Selection: A feature selection method based on mutual information is used to identify the most relevant and least redundant features for classification.
  • Classification: A Gaussian Mixture Model (GMM) is used as the classifier, which is well-suited for modeling multi-dimensional data distributions.
  • Correction Algorithm: To improve accuracy in real-world scenarios, a correction algorithm is implemented. This algorithm maintains the current activity classification if the new recognition result is the same as the previous one, or if the new result differs but the previous result was "staying" and has been consistent for a certain duration (e.g., 3 minutes). This helps to filter out transient misclassifications.
  • Battery-Saving Algorithm: To address the power consumption of continuous sensor usage, a dynamic algorithm is proposed. This algorithm intelligently turns sensors on/off or adjusts their sampling frequency based on the detected activity and speed (derived from GPS). For instance, the accelerometer can be turned off when the speed indicates bus or subway travel, and GPS reception cycles can be extended when not crucial for recognition.

Datasets

The study collected data using Samsung Galaxy S3 smartphones. The data was collected from ten subjects (two females, eight males, aged 24-28) carrying smartphones in four common positions: shirt pocket, pants pocket, bag, and hand. The collected activities included staying, walking, jogging, riding in a bus, and riding in a subway. The total length of collected datasets is 93,510 seconds. The data was distributed for training (80%), validation, and testing (20%).

Key Insights and Ideas

  • Vehicle-Specific Vibrations as Discriminators: The core insight is that the natural vibrations of vehicles like buses and subways, when analyzed in the frequency domain, provide unique signatures that can reliably distinguish them, even when user movement is minimal.
  • Orientation Independence is Crucial for Smartphones: The paper highlights that smartphone AR must account for varying phone orientations and positions, which is achieved through gyroscope-assisted gravity removal.
  • Correction Algorithm for Robustness: The correction algorithm is vital for real-world performance, preventing short-lived misclassifications from overriding the actual activity, especially in scenarios with intermittent sensor data or stationary periods within a vehicle.
  • Battery Saving Without Sacrificing Accuracy: The dynamic sensor management approach demonstrates that significant battery savings can be achieved without compromising recognition accuracy, making the system more practical for all-day use.

Key Results

  • Distinguishing Bus and Subway: The proposed method achieved an average accuracy of 91.82% using 44 features, including natural vibration features. This is a significant improvement over methods relying solely on accelerometer magnitude, which struggled to differentiate buses and subways.
  • Accuracy for Dominant Activities: The system achieved high accuracy for "staying" (94.10%), which is a critical activity in daily life.
  • Reduced Confusion: The confusion between bus and subway was reduced to a very low 1.2%.
  • Battery Efficiency: The battery-saving algorithm demonstrated significant energy conservation. For example, on a bus, battery usage decreased by about 55% over 35 minutes, and for other activities, efficiency increased by up to 22% over 30 minutes, without affecting recognition accuracy.
  • Free-Living Conditions: Experiments in free-living conditions showed good results, with only 3 out of 48 attempts being misrecognized over the entire activity cycle, highlighting the effectiveness of the correction algorithm.

Limitations and Future Work

  • GPS Dependency and Accuracy: The system relies on GPS for distinguishing between buses and subways, which can be problematic in areas with poor GPS reception or in underground subway sections. The accuracy of GPS data can also lead to misinterpretations of speed.
  • Subway Walking Confusion: A limitation identified is the potential for confusion when a user walks within a subway car, which might be misrecognized as "walking" initially before being corrected to "subway" after a period.
  • 3-Minute Interval for "Staying": The 3-minute interval used in the correction algorithm to confirm "staying" can lead to a delay in recognizing the actual activity, particularly when a user disembarks from a vehicle. This is acknowledged as a trade-off to prevent substantial errors.
  • Limited Activities: The current system is limited to recognizing a specific set of activities.
  • Future Work: The authors suggest expanding the number of recognizable activities by incorporating additional sensors like smartwatches. They also aim to improve the system's applicability in different countries by allowing for the input of different GPS beacon addresses. Further research will focus on addressing the limitations related to GPS signal reception and improving the distinction between walking within a subway and actual walking.

Summary updated

BibTeX
@misc{hur2017smartphone,
  author = {Tae Ho Hur and Jae Hun Bang and Dohyeong Kim and Oresti Banos and Sungyoung Lee},
  title = {{Smartphone Location-Independent Physical Activity Recognition Based on Transportation Natural Vibration Analysis}},
  year = {2017},
  doi = {10.3390/s17040931},
}
Figure