Abstract: In view of the shortcomings of current human behavior recognition algorithms,such as the poor timeliness of two stream convolutional neural network,the large number of parameters of 3D convolutional neural network and the high complexity of the algorithm,a space-time fusion pseudo-3D convolutional neural network model TFP3D was proposed based on 3D convolutional network and temporal fusion network.The 3D convolution splitting was used to reduce the large number of parameters brought by 3D convolution kernels.The temporal fusion module was added to ensure the effective transmission of the spatio-temporal features of human behavior information.Finally,the Kinetics dataset was used to pre-train the deep model to improve network speed while maintaining accuracy.A lot of experimental analyses were carried out on the common human behavior recognition dataset UCFl01,and the recognition results were compared with the current popular algorithms.The results showed that the proposed TFP3D designed was better than other methods,and the average recognition rate was greatly improved compared with other methods.

Key words: TFP3D network, temporal fusion network, pre-train, behavior recognition, deep learning

摘要： 针对当前人体行为识别算法中双流卷积神经网络时效性差、3D卷积神经网络参数多、算法的复杂度高等不足,提出了基于3D卷积网络和时空融合网络的时空融合伪3D卷积神经网络模型TFP3D。首先,使用3D卷积拆分减少3D卷积核带来的庞大参数量;其次,增加时空融合模块T-Fusion,保证人体行为信息时空特征的有效传递;最后,使用Kinetics数据集对深层模型进行预训练,在保证准确率的前提下提升网络速率。在常见的人体行为识别数据集UCFl01上进行了大量的实验分析,并将识别的结果和当前流行的算法进行比较,结果证明所设计的TFP3D优于其他方法,平均识别率相比其他方法有较大的提高。

关键词: TFP3D网络, 时间融合网络, 预训练, 行为识别, 深度学习