基于ResC3D神经网络的视频动作识别算法

Video Motion Recognition Algorithm Based on ResC3D Neural Network

基于深度学习的C3D网络被广泛应用于人体动作识别领域,但其识别准确率仍有较大提升空间。为了获得更好的识别准确率,本文改进了C3D的网络卷积结构并加深了网络层次,但网络深度的提高也会使网络出现梯度消失、梯度爆炸以及网络退化问题。针对这些问题,本文又将改进的C3D网络与ResNet50网络进行融合,在C3D网络中加入了残差结构,以此设计了ResC3D神经网络。将ResC3D网络模型在UCF101数据集上进行训练和验证,识别准确率达到了88.9%,相较于原始C3D网络的准确率提升了8.7%。

The C3D network based on deep learning is widely used in the field of human motion recognition,but its accuracy still has great room for improvement.In order to obtain better recognition accuracy,this paper improves the convolution structure of C3D network and deepens the network level,but the increase of network depth will also cause the network to have the problems of vanishing gradient,gradient explosion and network degradation.In order to solve these problems,the paper fused the improved C3D network and ResNet50 network,added the residual structure to the C3D network,and designed the ResC3D neural network.The ResC3D network model is trained and verified on UCF101 dataset,the accuracy rate is 88.9%,which is 8.7%higher than the original C3D network.