MFECNet: Multifeature Fusion Extra Convolutional Network Based on FMCW Radar for Human Activity Recognition

Abstract

The advent of the Internet of Things (IoT) has opened up a plethora of possibilities for human activity recognition (HAR) in a multitude of domains, including smart homes and health monitoring. However, conventional techniques, such as video and optical sensors, are constrained by shortcomings pertaining to privacy protection and environmental adaptation. Frequency modulated continuous wave (FMCW) radar has emerged as a prominent area of research due to its robust anti-interference capabilities and penetration, particularly its exceptional privacy protection. Nevertheless, there is a paucity of research in the field of computationally constrained mobile devices. Furthermore, the majority of the existing studies are limited by the use of a single input feature and similar activities that are susceptible to confusion. Consequently, the practical applications of the model must strike a balance between lightweight and accuracy. In this article, a self-built FMCW radar human activity dataset comprising seven classes of activities is built, and we have conducted a targeted study on one of the hazardous activities, falling. To address the existing problem, a threshold-convolutional denoising (TCD) algorithm for the generation of feature maps, with the objective of reducing the computational cost of the system, a multifeature fusion extra convolutional neural network (MFECNet) for activity recognition is proposed. In contrast to preceding the models of HAR systems based on FMCW radar, MFECNet combines the extra convolutional attention module and the universal inverted bottleneck (UIB) structure to develop a lightweight model. Introducing range-time maps based on the Doppler-time maps, enables the realization of multifeature input and recognizes the fused features, thereby enhancing the accuracy of the recognition of confusable activities. The resulting overall accuracy is 99.67%, in which the rate of missed and false alarms of falls was reduced to 0%. Meanwhile, the MFECNet validates the generalization of the model in the Glasgow dataset with a validation set accuracy of 99.62%, which is an improvement of 1.62%–3.62% compared to the model using the same dataset in references. The results show that the model proposed in this article achieves lightweight while improving the accuracy, which is more suitable for practical application scenarios.