Radar systems are increasingly being employed in healthcare applications for human activity recognition due to their advantages in terms of privacy, contactless sensing, and insensitivity to lighting conditions. The proposed classification algorithms are however often complex, focusing on a single domain of radar, and requiring significant computational resources that prevent their deployment in embedded platforms which often have limited memory and computational resources. To address this issue, we present an adaptive magnitude thresholding approach for highlighting the region of interest in the multi-domain micro-Doppler signatures. The region of interest is beneficial to extract salient features, meanwhile it ensures the simplicity of calculations with less computational cost. The results for the proposed approach show an accuracy of up to 93.1% for six activities, outperforming state-of-the-art deep learning methods on the same dataset with an over tenfold reduction in both training time and memory footprint, and a twofold reduction in inference time compared to a series of deep learning implementations. These results can help bridge the gap toward embedded platform deployment.

Radar is now widely used in human activity classification because of its contactless sensing capabilities, robustness to light conditions and privacy preservation compared to plain optical images. It has great value in elderly care, monitoring accidental falls and abnormal behaviours. Monostatic radar suffers from degradation in performance with varying aspect angles with respect to the target. Bistatic radar may offer a solution to this problem but finding the right geometry can be quite resource-intensive. We propose a bespoke simulation framework to test the radar geometry for human activity recognition. First, the analysis focuses on the monostatic radar model based on the Doppler effect in radar. We analyse the spectrogram of different motions by Short-time Fourier analysis (STFT), and then the classification data set was built for feature extraction and classification. The results show that the monostatic radar system has the highest accuracy, up to 98.17%. So, a bistatic radar model with separate transmitter and receiver was established in the experiment, and results show that bistatic radar with specific geometry configuration (CB2.5) not only has higher classification accuracy than monostatic radar in each aspect angle but also can recognise the object in a wider angle range. After training and fusing the data of all angles, it is found that the accuracy, sensitivity, and specificities of CB2.5 have 2.2%, 7.7% and 1.5% improvement compared with monostatic radar.

Population ageing has become a severe problem worldwide. Human activity recognition (HAR) can play an important role to provide the elders with in-time healthcare. With the advantages of environmental insensitivity, contactless sensing and privacy protection, radar has been widely used for human activity detection. The micro-Doppler signatures (spectrograms) contain much information about human motion and are often applied in HAR. However, spectrograms only interpret magnitude information, resulting in suboptimal performances. We propose a radar-based HAR system using deep learning techniques. The data applied came from the open dataset “Radar signatures of human activities” created by the University of Glasgow. A new type of hybrid map was proposed, which concatenated the spectrograms amplitude and phase. After cropping the hybrid maps to focus on useful information, a convolutional neural network (CNN) based on LeNet-5 was designed for feature extraction and classification. In addition, the idea of transfer learning was applied for radar-based HAR to evaluate the classification performance of a pre-trained network. For this, GoogLeNet was taken and trained on the newly-produced hybrid maps. These initial results showed that the LeNet-5 CNN using only the spectrograms obtained an accuracy of 80.5%, while using the hybrid maps reached an accuracy of 84.3%, increasing by 3.8%. The classification result of transfer learning using GoogLeNet was 86.0%.