Aggressive Action Recognition Using 3D CNN Architectures

This paper discusses an application of the transfer learning approach concerning human aggressive actions recognition task in video content. Comparative analysis of this approach was performed using various three-dimensional convolutional network architectures (3D CNN): Convolutional 3D Neural Network (C3D), Inception 3D Neural Network (I3D), Residual 3D Neural Network (R3D) based only on RGB frames processing. These 3D CNNs have trained on a composite aggressive action video dataset, that includes benchmark datasets for aggression recognition. The neural networks were evaluated in terms of accuracy, precision, recall, f1-score and loss function values metrics. The aggressive action recognition transfer learning approach using 3D CNNs showed impressive results on the considered metrics. Moreover, learning time in context of this approach was relatively short.