Human Behavior Recognition Based on IC3D

Aiming at the low accuracy of human behavior in videos, this paper proposes a method based on IC3D to identify human behavior in 3D-CNN neural network. A continuous 7-frame video images as input, using the gray level of video image, x-directional gradient, y-direction gradient, x-directional optical flow, y-directional optical flow, multichannel processing, after the 3D convolution and 3D pooling to extract the feature information with space and time dimensions. At the end of the network structure, the average pooling is processed to replace a layer of fully connected layer in the network structure, and the whole network is made regularization to prevent over-fitting, also because of the reduction of a layer of fully connected layer, so that the parameters required for training model are reduced greatly and the algorithm runs faster. Finally the action classification results are gotten through the fully connection layer and softmax classifier. Compared with the C3D, C3D+SVM and C3D+IDT+SVM three algorithms in UCF101 dataset, it is proved that the IC3D network framework has higher recognition accuracy and faster operation speed in the video human behavior recognition problem.