Human Activity Recognition Based on CWDCGAN-GP Data Augmentation and Improvement of Lightweight Network

In recent years, human activity recognition (HAR) techniques based on radar sensors and deep learning have shown great potential in areas such as fall detection, human-computer interaction, and security monitoring. However, building effective HAR datasets faces a number of challenges, including massive data collection and processing costs, as well as the problem of how to ensure sufficient data volume to avoid model overfitting. To address these issues, data augmentation using GANs has emerged as a solution. In this paper, we propose an improved data augmentation method, which is Conditional Wasserstein Deep Convolutional Generative Adversarial Network with Gradient Penalty (CWDCGAN-GP), to effectively solve the problems of training difficulty and low quality of generated samples in existing GANs. In addition, in order to adapt to the application requirements of mobile devices, we adopt the lightweight MobileNet-v2 network and introduce the coordinate attention (CA) and pyramid split attention (PSA) modules, which significantly improves the feature extraction and representation capabilities of the network. Finally, extensive experiments on three publicly available datasets show that the use of the CWDCGAN-GP model proposed in this paper not only achieves more stable training results, but also generates samples of significantly better quality than existing GAN methods, and this improvement effectively enhances the accuracy and generalization performance of the classification network. The proposed improved MobileNet-V2 network not only outperforms the existing action recognition algorithms in terms of average recognition accuracy and network complexity, but also shows excellent generalization performance, which is able to maintain stable and excellent performance in complex and diverse scenes.