基于迁移学习的危险行为识别方法研究

深度学习中卷积神经网络在行为识别领域有着良好的识别效果,但是由于深度学习需要较大数据集训练模型,而现今公开数据集中危险行为识别相关方向没有大量数据集。针对危险行为识别领域样本少、无法进行深度学习训练等问题,建立了危险行为识别数据集,并采用迁移学习方法对C3D网络模型进行迁移训练。结果表明,迁移学习后C3D网络模型对危险行为识别数据集平均识别率达到了83. 2%,可以有效识别危险行为动作。

Research on Dangerous Behavior Identificati on Method Based on Transfer Learning

The convolutional neural network in deep learning performs excellently in the field of behavior recognition, and a large data set is necessary for deep learning to train the model. However, there is no large data set related with dangerous behavior recognition in the open data sets at the moment. Aiming at the problems such as lacking samples in the field of dangerous behavior recognition and being unable to carry out training in deep learning, this thesis establishes the dangerous behavior recognition data set and adopts the transfer learning method to carry out transfer training for the C3D network model. It turns out that, after transfer learning, the average recognition rate of the C3D network model on the dangerous behavior recognition data set is up to 83.2%, which means C3D model can effectively recognize dangerous behaviors.