A New Double-flow Network via Decision Tree for Human Action Recognition

Traditional convolutional neural networks tend to lose time flow information when extracting video action features. The feature extraction ability of the novel model is reduced. To solve the above problems, first of all, RGB images in each video and their corresponding vertical dense optical flow sequence map are extracted to obtain the spatial and temporal information of the actions in the video. They input the spatial domain network and the temporal domain network respectively for pre-training, and then perform feature extraction after pre-training. Then, the parallel fusion strategy is implemented for the feature vectors with the same dimension. This process can enhance the representation ability of the feature vectors. Finally, the fused feature vectors are input into the decision tree for training and classification. Experiments on public datasets show that the new proposed method has lower error rate and high recognition rate. Human behavior analysis technology is of great significance to the current theoretical research and development. And how to make the algorithm applicable to different scenes, more accurate detection and improve the fall detection effect in complex environment has become a problem to be solved.