FC_C3D: A human behavior recognition method based on 3D fully convolutional network

Focusing on the problem of computational resources and storage space limitations faced by C3D convolutional neural network models when deployed in edge computing devices, this paper proposes a lightweight behavior recognition method for FC_C3D networks. Firstly, a lightweight 3D convolution scheme is proposed to reduce the model parameters by optimizing the number of 3D convolutional layers and introducing a batch normalization layer. Secondly, a network structure in full convolutional form is designed with adaptive mean pooling and convolutional classification to significantly reduce the computational resource requirements. Finally, the Adam gradient descent optimizer and Mish activation function are used to accelerate model convergence and improve accuracy. The experimental results demonstrate that the accuracy of this paper's method on the UCF101 dataset as well as the HMDB51 dataset is improved by 39.6% and 36.7%, respectively, the number of parameters is reduced by 85%, and the operation speed is improved by 18.1%.