The Pre-impact Fall Detection using the Quantization based on ResNet

Abstract

Purpose Fall is the one of the major health risks for the older people.  Hip fractures are considered the most dangerous among various injuries, as they can reduce mobility and lead to many complications, and even death (Hagen, G et al., 2020). Some researchers tried to protect the user using wearable airbag based on the threshold-based algorithm (Jung, H et al., 2021). The more accurate algorithm was developed using the deep learning methods, but it was heavy and needed GPU or PC (Yu et al., 2021). This study was focused on developing the lightweight deep learning algorithm to detect pre-impact falls for targeting edge devices. Method The KFall public dataset was used in this study. The experimental protocol consisted of 15 fall movements and 21 activities of daily living (ADLs). The 9-axis IMU sensor data with 100 Hz sampling frequency were acquired from the waists of 32 young male subjects. Data of 26 subjects were used to train the deep learning models and those of remained 6 subjects were used to test the models. We developed the fall detection algorithm based on ResNet method. It consisted of 2 convolution layers, 4 convolution block, 3 identity block, average pooling layer, and SoftMax layer. We removed one by one except for the factors that affect the output size (figure 1). The ResNet14 model was transformed to a flat buffer type and two quantization techniques were applied using TensorFlow Lite. Results and Discussion Table 1 showed the performance of ResNet algorithms. When continuously reducing the layers, demanded memory decreased, but there was no dramatic change in accuracy. It suggested that our model was sufficiently deep and additional identity blocks were not needed. Table 2 showed the performance of ResNet14 models according to quantization. The integer quantization showed the smallest size of memory but the worst accuracy. The float16 quantization showed bigger size of memory than the integer quantization, but the accuracy maintained. As a result, the float16 quantization showed 98% of accuracy with 539±278 ms of sufficient lead time and 0.104 MB of memory size.

References

Hagen, G., Magnussen, J., Tell, G., & Omsland, T. (2020). Estimating the future burden of hip fractures in Norway. A NOREPOS study. Bone, 131, 115156.

Jung, H., Koo, B., Kim, J., Kim, T., Nam, Y., & Kim, Y. (2020). Enhanced algorithm for the detection of preimpact fall for wearable airbags. Sensors, 20(5), 1277.

Yu, X., Jang, J., & Xiong, S. (2021). A large-scale open motion dataset (KFall) and benchmark algorithms for detecting pre-impact fall of the elderly using wearable inertial sensors. Frontiers in Aging Neuroscience, 399.