Figure 1.

An overview of our proposed method. (a) The input of the convolutional neural network (CNN) is the original time series data. (b) We use CNN to train a high-accuracy pre-impact fall detection model. (c) Based on the class activation mapping (CAM) method, we obtain the hot map of the original time series, which highlights the contributing region. By analysis on this region in detail, we manually extract effective features and characteristics. (d) Combining threshold detection on the obtained feature region, the method can be applied to wearable devices with comparable accuracy.