An Effective Action Recognition Method Based on Image Coding and a Dual-Channel Fusion Network

Abstract

Action recognition is a research hotspot in artificial intelligence, with significant applications in intelligent sports analysis, health monitoring, and human–computer interaction. Traditional methods rely on high-frame-rate cameras or complex motion capture systems, which are costly and highly dependent on environmental conditions. In contrast, data-driven methods based on wearable sensors have gained widespread attention due to their portability and cost-effectiveness. In this article, we propose an action recognition method based on image encoding and a dual-channel feature extraction network. We convert time-series data collected from wearable sensors into color images through image encoding, fully preserving the temporal information and multidimensional feature relationships in the data. Then, we design a dual-channel feature extraction network that extracts complex features using a multiscale spatial channel attention (MSCA) module, a dual-stream alternating feature fusion (DAF) module, and a weighted loss function (WFL). We conducted experiments on the USC-HAD and PAMAP2 datasets, demonstrating that our method outperforms several state-of-the-art methods. Ablation studies further verify the contributions of the backbone network, fusion module, classifier, and loss function to the overall performance. Overall, our method provides a new solution for action recognition tasks and shows broad application prospects.