Lightweight and Generalizable Multi-Sensor Human Activity Recognition via Cascaded Fusion and Style-Augmented Decomposition

A team of researchers has developed a novel AI framework designed to make human activity recognition on wearable devices significantly more efficient and robust. The work, led by Wang Chenglong, Zhuo Yan, Ding Wenbo, and Chen Xinlei, tackles a core challenge in Wearable Human Activity Recognition (WHAR): balancing accuracy with the computational limits of small devices. Their solution retains the common 'decomposition-extraction-fusion' structure but innovates with two key components to replace bulky and brittle parts of existing models.

First, they introduced a Cascaded Fusion Block (CFB) to handle the complex task of fusing data from multiple sensors (like accelerometers and gyroscopes). This block uses a process of 'compression-recursion-concatenation-fusion' to achieve efficient feature interaction, completely avoiding the computationally expensive attention mechanisms used in many state-of-the-art models. Second, they integrated a MixStyle-based data augmentation module before feature extraction stages. This technique subtly mixes the statistical styles (mean and variance) of different data samples in a training batch, artificially creating diverse data scenarios. This 'style-augmented decomposition' enhances the model's ability to generalize to new users or environments without altering the core activity data.

The results are compelling. Evaluated on standard benchmark datasets Realdisp and Skoda, the proposed framework not only achieved higher accuracy and macro-F1 scores than existing methods but did so while reducing computational overhead by more than 30%. This dramatic efficiency gain comes without sacrificing the model's ability to understand intricate spatio-temporal relationships within multi-sensor time-series data. The framework meticulously maintains independence at the sensor, variable, and channel levels during its initial decomposition phase, leading to more robust and interpretable feature extraction in the subsequent fusion steps.