This paper shows a practical 60 GHz-band radar-based method to recognize behaviors for bed monitoring systems. Time-range images, derived from short-term data, were produced through the processing of frequency-modulated continuous-wave (FMCW) radar signals. Using the generated images, behaviors such as leaving the bed, lying on the bed, and sitting on the bed were accurately classified with a classification accuracy of 96% using a convolutional neural network with a relatively light-sized model of MobileNet. Furthermore, six representative behaviors in bed monitoring were classified with an accuracy of 83.1%. These findings indicate that 60 GHz-band millimeter-wave radar holds promise as a non-invasive bed monitoring tool.
In the elderly care sector, the shortage of careers relative to the number of individuals requiring
assistance has led to an increased burden on each caregiver. Consequently, round-the-clock
monitoring in nursing homes to prevent falls or wandering is required and is contributing to
caregiver workload. Previous researches have focused on sensor-based monitoring systems to
track bed-bound behaviors, with early detection of bed-leaving being crucial for identifying
abnormal situations. These systems have employed various devices such as cameras, pressure
sensors, wearable sensors, and infrared cameras [
To tackle the above problems, recent research has explored the integration of radar technology
into monitoring systems designed for the elderly [
Thus, we developed a bed-leaving detection method using short-term radar data [
In this paper, we report the experimental results for multinomial classification of human behaviors for bed monitoring. We assume six classes of behaviors related to bed-leaving and bed-lying activities. Classification accuracies using practical CNN models with the inputs of the short-term time-range images were evaluated for 10 participant data.
Figure 1 shows the experimental setup. A 60 GHz FMCW radar was installed at a height of 40 cm from the surface of the bed. The radar bandwidth was B = 6.8 GHz, the range resolution was ∆ R = c/2B = 2.4 cm (where c is the speed of light) and its frame rate was 80 Hz. Radar directivity in a plane parallel to the bed surface was 120°.
The participants were ten young adults. Instructions were given before measurement and each behavior was measured for the 60 s with the pulse repetition interval of 80 Hz. Our experiments aim to classify representative behaviors on and around the bed via the unstrained measurements using the radar. The assuming behaviors were "bed-leaving", "bed-lying", "sitting square", "long sitting", "standing outside" and "lying outside". Fig. 2 shows the experimental sites for some classes. Each behavior is defined as follows.
• bed-leaving: A participant is not in the radar measurement area. • bed-lying: A participant is lying on the bed. This includes the participant’s slight motions in the bed. • sitting square: A participant is sitting on the edge of the bed. • long sitting: A participant is sitting with extended legs on the bed. • standing outside: A participant is standing by the bed. • lysing outside: A participant is lying on the floor by the bed. This class assumes abnormal situation.
To classify the behaviors, we calculate the time-range distributions from the radar received
signals similar to our previous work [
Fig. 3 shows the examples of time-range images of the six classes. In these images, the vertical axis means that the lower part of the image is further away. There are no significant components in the "bed-leaving" image because the participant does not exist in the measurement area. The "lying outside" image is similar to "bed-leaving" except for the slight components in a relatively far range. We can confirm the similar significant components in the "bed-lying", "sitting square", and "long sitting" images. In the "standing outside" images, the significant components are confirmed in far range compared with "bed lying". We used these images for the CNN inputs to classify the six classes based on the slight diferences described above.
In this section, the classification results for representative behaviors related to bed-leaving and -lying movements are presented. For each classification, the data to evaluate classification performance was obtained from 10 participants, and the mean classification accuracy in the leave-one-subject-out was evaluated. We generated about 100 time-range images corresponding to each behavior for all participants.
The behavior corresponding to each image is classified using the CNN. For the CNN architecture, we used the ResNet-18 [8] because it was demonstrated to be eficient for radar-based human activity recognition problems [9]. This study used the ResNet-18. In addition, we also used MobileNetV3 [10] because the ResNet model size is relatively large, and a smaller model size is required for practical use based on edge computing. For both models, we performed training for 50 epochs and used a batch size of 64. The learning rate was 0.01 and was decreased by multiplying it by 0.5 every ten epochs. These hyperparameters were empirically optimized.
Because the detection of bed-leaving is the most important function for daily bed monitoring systems, this subsection considers the classification of bed-leaving and other five behaviors. Table 1 shows the results for the ResNet with a mean classification accuracy of 97.8%. Suficient accuracy was achieved, and the rate of misclassification of bed-leaving as other behaviors is lower. The result is considered safe in terms of developing a monitoring system. Table 2 shows the results for the MobileNet with a mean classification accuracy of 95.9%. Although the accuracy of MobileNet is lower than that of ResNet, suficiently accurate detection of bed-leaving is achieved even when using MobileNet with a smaller model size.
We classify six behaviors to develop more useful multinominal behavior classification functions for the bed monitoring systems. Table 3 shows the results for the ResNet, with a mean classification accuracy of 89.7%. As indicated in this table, the classes "sitting square" and "long sitting" were misclassified because the features of these classes were not suficiently expressed in the time-range image. In addition, "lying outside" was also largely misclassified because the part of the participant was out of measurement area, and it concluded time range images included less information. However, the important classes of "bed-leaving" and "bed-lying" were accurately classified. Thus, we achieve the behavior classifications with moderate accuracy even using the images generated from short-term data of 0.5 s.
Table 4 shows the results for the MobileNet with the mean classification accuracy of 68.5%. Because MobileNet is a simple CNN model, classification accuracy is lower than that of ResNet. However, the model sizes of the ResNet and MobileNet in this study were 18.0 and 0.2 MB, respectively. Thus, the performance improvement using the MobileNet is important future study to develop practical bed monitoring systems with simple implementation. Multiple images corresponding to multiple receiving antennas can be used to improve the classification performance.
To develop practical remote bed monitoring systems for nursing care, this study aimed to classify various behaviors related to bed-leaving and bed-lying using short-term data generated via the received signals of 60 GHz millimeter-wave FMCW radar. The CNN, utilizing time-range images generated from radar data at 0.5 s intervals, achieved the accurate classification of bed-leaving and other representative behaviors (bed-lying, sitting square, long sitting, standing-out of the bed, and lying-outside of the bed). Even when we used the simple MobileNet, the detection rate of bed-leaving was 95.9%.In addition, we achieved the above six types of behaviors with 89.7% accuracy using the ResNet. However, the classification accuracy using the MobileNet was 68.5%, although it has merits in suficiently smaller model size. Thus, to develop practical bed monitoring systems, future investigation to improve the classification accuracy using the multiple images of multiple receivers is required. related to bed-leaving and bed-lying using millimeter-wave fmcw radar., in: ATAIT, 2023, pp. 105–114. [8] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778. [9] S. Z. Gurbuz, M. G. Amin, Radar-based human-motion recognition with deep learning: Promising applications for indoor monitoring, IEEE Signal Processing Magazine 36 (2019) 16–28. [10] A. G. Howard, Mo-bilenets: Eficient convolutional neural networks for mo-bile vision applications, arXiv preprint arXiv:1704.04861 (2017).