The purpose of pre-impact fall detection systems is to detect the occurrence of falls during falling and support fall protection systems for preventing severe injuries. This repo has implemented a pytorch-based pre-impact fall detection systems via CNN-ViT knowledge distillation, namely PreFallKD, to strike a balance between detection performance and computational complexity. The proposed PreFallKD transfers the detection knowledge from the pre-trained teacher model (Vision Transformer) to the student model (lightweight convolutional neural networks). Additionally, we apply data augmentation techniques to tackle issues of data imbalance. We conduct the experiment on the KFall public dataset and compare PreFallKD with other state-of-the-art models.
The input window is the 50 frames IMU signals, which include triaxial acceleration data, triaxial gyroscope data, and triaxial Euler angle data. The ViT-tiny is the teacher model and the lightweight CNN is the student model. The student model can learn the high dimension knowledge from the teacher model by Kullback-Leibler divergence loss function (KL Divergence Loss) and learn ground truth by Focal loss.
python == 3.9
pytorch >= 1.12.1
numpy >= 1.23.2
matplotlib >= 3.5.3
pandas >= 1.4.3
einops >= 0.4.1
This is still a work in progress and is far from being perfect: if you find any bug please don't hesitate to open an issue.