Traditional human action recognition methods often face limitations such as low accuracy, high model complexity, and challenges in practical deployment. To overcome these issues, this paper explores integrating sensor principles, digital signal processing, and deep learning techniques, aiming to enhance recognition performance while maintaining computational efficiency. The goal is to develop a robust, lightweight recognition framework suitable for real-time use in edge computing environments like smart homes, elder care, and security monitoring.

This paper introduces a novel human action recognition approach based on a dual-stream, cross-fusion, and bi-level programming model that incorporates a convolutional block attention module and Transformer (CBAM-Transformer). A frequency-modulated continuous wave radar platform is built to collect human motion data. The raw radar signals undergo preprocessing, including dimension reorganization, clutter suppression with a Butterworth filter, and time-frequency analysis, resulting in a dual-mode dataset with range and micro-Doppler spectrograms. The network architecture features a bi-level design: the upper layer uses a dual-stream CBAM-Transformer network to facilitate multi-domain feature interaction and global dependency modeling, while the lower layer employs a lightweight MobileNetV3 network. Knowledge distillation helps transfer representational power from the upper to the lower network, enabling model compression and faster inference. The model is trained using stochastic gradient descent with momentum and a cosine annealing learning rate.

Testing on a self-collected dataset of 4000 samples across eight action categories, the proposed method achieves a recognition accuracy of 96.73%. It also significantly reduces computational load, with only 5.1 million parameters and 1.45 TFLOPs of floating-point operations. Ablation studies confirm the contributions of each component: dual-stream input outperforms single-stream models; the progressive addition of cross-fusion, Transformer, and CBAM modules improves recognition accuracy. Comparative experiments show that the bi-level model maintains high accuracy while drastically cutting resource needs compared to the upper-layer network alone, striking a balance between accuracy and efficiency.

The CBAM-Transformer-based dual-stream, cross-fusion, and bi-level model effectively tackles challenges in human action recognition, such as data scarcity, limited feature representation, and high computational cost. This approach achieves notable improvements in recognition accuracy while remaining lightweight for real-time deployment in resource-constrained edge environments. The paper highlights the advantages of combining attention mechanisms, cross-modal fusion, and bi-level optimization for developing efficient and reliable action recognition systems, offering a practical, scalable solution for intelligent perception tasks.