A human action recognition method based on spatiotemporal information interaction
Tian Wei · Advances in Engineering Innovation · 2025
In the field of deep learning, current human action recognition algorithms often treat temporal information, spatial information, and background information equally, which leads to limited recognition accuracy. To address this issue, this paper proposes a human action recognition algorithm based on spatiotemporal information interaction. First, a dual-pathway network is proposed to learn spatial and temporal information at different refresh rates. The network includes a sparse pathway operating at a low frame rate to capture spatial semantic information, and a parallel dense pathway operating at a high frame rate to capture temporal motion information. Second, to extract more discriminative features from videos, a cross-dual attention interaction model is introduced to focus on key regions of video segments and explicitly exchange spatiotemporal information between the two pathways. Experimental results show that the proposed algorithm achieves recognition accuracies of 97.6% on the UCF101 dataset and 78.4% on the HMDB51 dataset, outperforming the novel SlowFast algorithm by 1.8% and 1.4%, respectively. Combined with a nighttime image enhancement algorithm based on MDIFE-Net curve estimation, the method achieved an accuracy of 83.2% on the ARID nighttime datasetan improvement of 22.9% over the performance before image enhancement. This demonstrates the methods strong potential for real-world nighttime action recognition applications.