Abstract
To study the rules and characteristics of the daily actions of the aged, a set of multi-sensor wearable device has been developed. As for the segmenting method of action data collected from the aged wearing this device, the human action sequence segmentation method has been proposed based on the motion characteristics in this paper. That is to say, collecting the data for Fourier transform, determining the equilibrium position of the motion, and then returning back to the time domain and determining the segmentation point by judging the changing trend of signal waveform, so as to realize the segmentation of the entire signal. Experiments show that this method can effectively segment the daily actions of the aged, extract valid action sequence, and is particularly effective in the segmentation of regular actions.
Introduction
As the global population ages, monitoring the daily activities and gait of the elderly has become a priority for healthcare and rehabilitation. While wearable sensors like accelerometers offer a low-cost, flexible way to collect this data, they produce a continuous stream of information that is difficult to analyze directly. To calculate metrics such as step length or frequency, researchers must first perform segmentation: breaking the long stream of data into individual, repeatable units. Traditionally, this is done manually, which is time-consuming and subjective. This paper proposes an automated method to slice continuous wearable sensor data into discrete "action cycles" using Fast Fourier Transform (FFT).
Methodology
The researchers developed a multi-sensor wearable system consisting of a 3D inertial sensor (accelerometer) at the waist and two flexible sensors placed on the knees to measure joint angles.
The core of the segmentation approach involves three main steps:
- Frequency Domain Analysis: A Fast Fourier Transform (FFT) is applied to the raw signal from the flexible sensors. In the frequency domain, the position with a frequency of 0 represents the "equilibrium position"—the baseline around which the joint angle fluctuates.
- Equilibrium Intersection: By mapping this equilibrium line back onto the time-domain signal, the algorithm identifies where the signal intersects the baseline.
- Cycle Definition: The space between two continuous intersections represents half a cycle; three continuous points define one full motion cycle. From here, the algorithm automatically calculates the peak, valley, amplitude, and duration of each movement.
Experimental Setup
The study involved a cohort of 8 testers (2 men, 6 women) from the Wuhan Baoxie Health Service Center, aged 60–70. The testers performed six specific actions: running, sitting, standing, sleeping, rising, and walking. Data was sampled at 10ms intervals, capturing 3D acceleration (X, Y, Z) and two channels of flexible sensor data (left/right knee curvature).
Results and Discussion
The authors found that flexible sensors on the knees provide the most optimal reference data. Their waveforms are smoother and more periodic compared to the noisy waist acceleration data, making them easier to segment.
The automatic segmentation was compared against manual segmentation using a Support Vector Machine (SVM) to recognize the actions. The results showed a clear advantage for the automated method:
| Method | Accuracy (Running) | Accuracy (Walking) |
|---|---|---|
| Manual Segmentation | 50% | 70% |
| Automatic (FFT-based) | 100% | 75% |
The dramatic jump in running accuracy suggests that for high-frequency, regular actions, the algorithm is nearly perfect at isolating clean cycles that the recognition model can easily classify.
Conclusion
The FFT-based segmentation method provides a robust, objective baseline for zero-crossing detection in gait analysis. While signal noise from elderly movements remains a challenge, the automated approach ensures every segment is sliced at the exact mathematical intersection, leading to more consistent data for machine learning models. Future work will focus on more sophisticated noise-filtering techniques.
Summary updated