摘要
针对遮挡环境下人体行为信息的不完整性,导致行为识别准确率低的问题,提出了一种改进的注意模型(IA-Net)。为减少参数剧增,降低计算消耗,采用自适应卷积(adaptive convolution)层代替压缩提取模块(SE-block)中的全连接(FC)层。同时为防止SE-block产生神经原失活的问题,在激活层(sigmoid)之前加批标准化(BN)层对数据进行标准化处理,使得输入给sigmoid激活函数之前数据处于该函数的非饱和区,提出改进的注意力模块(ISE-block)。将ISE-block嵌入到残差网络ResNet50中,形成ISE-ResNet50网络,用于提取人体行为特征,提升重要特征权重同时抑制非重要特征权重。考虑复杂行为需长时间序列表示其前后动作依赖关系并突出主要特征,将ISE-ResNet50网络的输出送给具有注意力机制的长短期模块(ATT-LSTM),最终形成IA-Net模型,实现端到端的行为识别。在HMDB51、UCF101两个数据集上进行实验,提出的IA-Net模型分别获得86.32%和97.78%的识别精度。与时空残差网络ST-ResNet在HMDB51数据集上的识别精度相比提升了1.52%。实验结果表明,IA-Net在行为识别方面具有更高精度。
An improved attention modal(IA-net)is proposed to solve the problem that the incompleteness of human behavior information in occlusion environment leads to the low accuracy of human behavior recognition.In order to reduce the sharp increase of parameters and reduce the calculation consumption,the adaptive convolution layer is used to replace the full connection(FC)layer in the compression extraction module(SE-block).Simultaneously,to prevent the problem of neuron inactivation in SE-block,a batch normalization(BN)layer is added before the activation layer(sigmoid)to standardize the data,so that the data before input to the sigmoid activation function is in the unsaturated region of the function,and an improved attention module(ISE-block)is proposed.ISE-block is embedded into the residual network ResNet50 to form ISE-ResNet50,which is used to extract human behavior features,improve the weight of important features and suppress the weight of non-important features.Considering that complex behaviors require a long-term sequence to represent the temporal dependencies of front and rear actions and highlight the main features,the output of the ISE-ResNet50 network is fed into the attention mechanism long and short-term module(ATT-LSTM),and finally the IA-Net modal is formed to realize End-to-end behavior recognition.Experiments are carried out on the HMDB51 and UCF101 datasets.The IA-Net method proposed in this paper is obtained 86.32% and 97.78% recognition accuracy respectively.Comparing with the spatiotemporal residual network ST-ResNet,the recognition accuracy on the HMDB51 dataset is improved by 1.52%.The experimental results show that the IA-Net proposed in this paper has higher accuracy in action recognition.
作者
张银环
Zhang Yinhuan(Weinan Vocational&Technical College,Civil&Architectural Engineering,Weinan 714000,China;chool of Mechatronic Engineering,Xi'an Technological University,Xi'an 710021,China)
出处
《国外电子测量技术》
北大核心
2022年第6期52-59,共8页
Foreign Electronic Measurement Technology
基金
国家自然科学基金(6207010855)
渭南市科学技术局(2020ZDYF-JCYJ-235)项目资助。
