基于多尺度3D卷积神经网络的行为识别方法

近年来卷积神经网络(convolutional neural network,CNN)在行为识别任务中取得了较大的进展.然而,现有的神经网络方法往往只注重高层语义信息的利用,对浅层特征信息挖掘利用不够.针对这一问题,提出一种基于3D卷积(convolution 3D,C3D)的多尺度3D卷积神经网络的行为识别方法.该方法受到特征金字塔结构的启发,在原C3D的基础上融合C3D的浅层特征信息,实现端到端的行为识别.同时该方法以现有的深度学习理论为基础,利用迁移学习的思想,将C3D和该方法中相同模块部分的参数迁移到本方法中,以降低模型的训练时间.通过在UCF101数据集上进行实验,实验结果表明,提出行为识别方法的分类精度达到84.56％,分类效果优于原C3D分类网络.

Multiscale 3D convolutional neural network for action recognition

In recent years, convolutional neural network (CNN) has made great progress in action recognition tasks. However, the existing neural network methods tend to focus on using high-level semantic information, but not enough on the shallow feature information mining. To solve this problem, a multi-scale 3D convolution neural network for action recognition based on 3D convolution (C3D) is proposed in this paper. Inspired by the feature pyramid strategy, this method combines the shallow feature information of C3D to realize the end-to-end action classification. At the same time, the method uses the existing deep learning theory and utilizes the idea of transfer learning to migrate the parameters of C3D to reduce the training time of the model. Experiments on UCF101 dataset show that the classification accuracy of the proposed action recognition method reaches 84.56%, and the classification result is better than that of the original C3D classification network.