Crowd Behavior Recognition Based on 3D Res-Inception Network Structure

For the task of crowd behavior recognition, combining traditional computer vision and deep learning, a two branch 3D Res-Inception network structure was proposed. Based on the two dimensional convolutional neural network, a spatiotemporal residual unit was designed to extract spatiotemporal features. Based on this, a two branch 3D Res-Inception structure was designed to fuse the appearance and motion features of the crowd, and mirroring and tailoring were employed to amplify the population of CUHK crowd dataset. The experimental results show that the method of data augmentation is suitable for dense crowd video recognition. Compared with the population descriptor method based on population transition, convolution neural network-long-short-term memory network (CNN-LSTM) and three dimensional convolution network (3D CNN), the recognition accuracy of the method with the proposed two branch 3D Res-Inception network structure is significantly improved, up to 95.48%.