| Abstrakt: |
In recent years, human activity recognition (HAR) using frequency module continuous wave (FMCW) radar is an effective tool that has been widely used in the fields of healthcare, smart driving, and smart living due to its convenience, inexpensiveness, and accuracy. Past studies have mainly investigated the improvement of the accuracy of HAR models while neglecting the deployment of the models. Therefore, we propose a model named FMCW lightweight vision transformer (FML-Vit) for HAR, primarily consisting of the FML-Vit block and FML-Vit subsample modules. The FML-Vit block, by incorporating a cascaded linear self-attention mechanism in place of the traditional multi-head attention mechanism, can transform the time complexity from ${O}\text {(} {k}^{{2}} \text {)}$ to ${O}\text {(}{k}\text {)}$ . The FML-Vit subsampling modules perform dimension reduction and feature reallocation, while the context broadcasting (CB) module is used to reduce the density in the original attention maps, thereby increasing both the capacity and generalizability of the ViT. The proposed algorithm is compared with nine different state-of-the-art methods on self-datasets and open-source datasets. The results demonstrate that FML-Vit outperforms other current lightweight networks with the fastest inference. |
