面向时空特征融合的GSTIN动作识别网络

GSTIN with spatiotemporal feature fusion for video action recognition

视频动作识别是计算机视觉领域一个十分具有挑战性的课题,主要任务是利用深度学习等视频智能分析技术识别的深层信息推导出视频人体行为动作。通过结合双流卷积神经网络和三维卷积神经网络的结构特点,提出了一种面向时空特征融合的GSTIN(GoogLeNet based on spatio-temporal intergration network)。GSTIN中设计了时空特征融合模块InBST(inception blend spatio-temporal feature),提升网络对空间特征与时间特征的利用能力;在时空特征融合模块InBST基础上,构建了适合动作识别的多流网络结构。GSTIN在动作识别数据集UCF101、HMDB51上识别精度分别达到了93.8%和70.6%,这表明GSTIN与其他动作识别网络相比具有较好的识别性能。

Video action recognition is a very challenging topic in the field of computer vision.The main task is to use the deep information by intelligent video analysis technology such as deep learning to recognize the human behavior.To further improve the performance of the two main frameworks,this paper proposed GSTIN for spatiotemporal feature fusion.GSTIN designed a spatiotemporal feature fusion module InBST,which could make network obtain the interactive temporal and spatial information.Based on the spatiotemporal feature fusion module InBST,GSTIN constructed a multi branch GSTIN suitable for action recognition.GSTIN was tested two classic video action recognition datasets UCF101 and HMDB51.Compared with the action recognition networks,experimental results show that GSTIN has better recognition performance.