Rescue workers at disaster sites and construction site operators routinely perform demanding tasks. They become fatigued due to these demanding tasks. This fatigue causes discomfort and a lack of motivation at work which afects both work eficiency and safety. Therefore, it is necessary to constantly monitor their physical condition. Several existing technologies are known to detect fatigue using a face image or wearable device. However, these systems have several issues. Recording the face carries a risk of invading the subject's privacy, and it is inconvenient to constantly wear a wearable device. In recent years, a head-mounted camera which is attached to a helmet is widespread. Therefore, a human-recorder has been proposed to obtain the user's biometric information and to detect fatigue from the head-mounted camera. The human-recorder extracts head sway in a video recorded by the head-mounted camera, and analyzes the time change of head sway. In this paper, the estimated pulse rate is calculated using the human-recorder to confirm that the pulse rate can be estimate in real time. Furthermore, the efect on the accuracy of the pulse rate estimation is verified by the camera frame rate. From the results, it was verified that the pulse rate can be estimated using the human-recorder. Also, it was confirmed that pulse rate estimation is possible regardless of the camera frame rate.
sites. Also, it can be used hands-free operation. Therefore, a human-recorder has been proposed to obtain the user’s biometric information for fatigue detection using the head-mounted camera [5, 6, 7]. The human-recorder extracts head sway in a video recorded by the head-mounted camera, and analyzes the time change of head sway. In this paper, the estimated pulse rate is calculated using the human-recorder to confirm that the pulse rate can be estimate in real time. Furthermore, the efect on the accuracy of the pulse rate estimation is verified by the camera frame rate. This analysis is important because it supports more practical and real-time use at disaster sites and construction site.
(
Head-mounted (
(
P time ( , )
Frequency
The feature coordinate point is calculated using optical flow to extract the user’s head sway (Fig 2 (b)). The first feature point 1(1, 1) is determined from the first frame in the recorded video. The feature point (, ) each frame is calculated by tracking the feature point 1(1, 1) using optical flow, and the time change of the feature coordinate point in x-coordinate (Fig 2 (c)) and y-coorfinate (Fig 2 (d)) are extracted. The x-coordinate time change is analyzed in this paper.
FFT Band-pass filter r e w o P
(c) Battery Case Pulse sensor
In this experiment, a video is recorded by the human-recorder and analyzed to confirm that real-time pulse rate estimation is possible. Figure 4 shows the overview of pulse rate estimation using human-recorder. Raspberry Pi Zero 2W controls camera module 3 at 50 fps in this prototype. Three participants put on this prototype and record a video for 120 seconds in a sitting position. Then, the estimated pulse rate is calculated by analyzing the video recorded using the human-recorder. Additionally, the actual pulse rate of the participants is measured using a pulse sensor while they record a video to compare with the estimated pulse rate. (a) Recording situation
(b) Human-recorder
In this experiment, the efect on the accuracy of the pulse rate estimation is verified by setting the fps values to 30, 40, 50, and 60 in the same situation as in Section 3.1. The Root Mean Square Error (RMSE) is used to evaluate the accuracy of predictions by calculating the square root of the mean of the squared diferences between predicted and actual values. It is defined by equation 3 provides a single value that summarizes how close the estimated values are to the true values. Smaller RMSE values indicate higher accuracy and better model performance. ⎯
RMSE = ⎷⎸⎸ 1 ∑︁ ( − ˆ )2
=1
(
In this paper, the estimated pulse rate was calculated using the human-recorder to confirm that pulse rate can be estimated in real time. Furthermore, the efect of the camera’s frame rate on the accuracy of pulse rate estimation was evaluated. The results demonstrated that the pulse rate measured by the human-recorder was in close agreement with the actual pulse rate. In other words, the human-recorder enables pulse rate estimation. Based on the RMSE results, it can also be confirmed that the diference in fps has little efect on the accuracy of pulse rate estimation. However, it was observed that the number of reliable feature coordinate points decreases as the fps decreases, which may negatively afect estimation accuracy. To address this issue, future work will focus on developing methods to interpolate missing or sparse feature point data, especially under low-fps conditions, in order to maintain stable performance.
The author(s) have not employed any Generative Al tools. “Verification of Fatigue Detection Human-recorder Using Depth Estimation Method,” , Vol. 2025-EMB-68, No. 30, pp. 1 – 6, March 2025. [8] https://www.raspberrypi.com/documentation/computers/raspberry-pi.html (Accessed 1 Ju ne 2025.) [9] https://www.raspberrypi.com/products/camera-module-3/ (Accessed 1 June 2025.)