Cross-spectral human behavior recognition based on deep convolutional networks for global temporal representation

The purpose of behavior recognition is to recognize the actions of the human body in action. It plays a great role in surveillance, video recommendation, and human-computer interaction with video. With the rise of neural networks, behavior recognition has also continued to develop and progress and has reached a relatively advanced level. However, behavior recognition is still insufficient in recognizing complex human movements and recognizing videos in different bands. To solve this problem, this paper establishes a convolutional neural network (CNN) cross-spectral human behavior recognition algorithm based on global time domain representation. It adopts the method of time-domain feature extraction, construction of optimized convolutional neural network layers, and global time-domain cross-spectrum construction. It also uses videos from the unified compliance framework (UCF)-sports and UCF-11 datasets for experiments. Experiments show that the algorithm achieves an average accuracy of 90% in the behavior recognition of UCF-sports. It still maintains an average accuracy rate of > 90 % in the more complex behavior recognition of UCF-11, and the highest accuracy rate is 93%.