Behaviour dataset using a K-band UWB linear FMCW radar

Current radar fall detection techniques based on deep learning (DL) networks are often too complex for real-time detection. This paper proposes a real-time fall detection approach by reducing the complexity of the DL networks and the UWB radar hardware requirements. A multi-indoor scene behaviour dataset of 40 subjects is established using K-band UWB radar. A sliding window-based dataflow augmentation method is proposed to augment and balance the given datasets. A simple radar signal preprocessing model that is suitable for hardware with low sampling rates provides appropriately sized images of range-time spectrograms for DL. A lightweight DL network is designed to realise real-time fall detection for radar-embedded devices. The dataset division process follows the principle of mutual exclusion. The five-fold crossvalidation results (F1-score = 0.9896 ± 0.0047) and testing results obtained in a new scene (F1-score = 0.9872) confirm that only using 2-second range-time spectrogram images as inputs of the proposed network is sufficient for achieving good classification performance and strong generalisation ability. The proposed network size is 1.9178 Mb, with of 37.9978 M floating point operations (FLOPs) when the input image size of the radar range-time spectrograms is set to 112×112. Compared with currently popular DL networks, the proposed network achieves a good compromise between classification performance and computational complexity.