Behavior Identification based on Improved Two-Stream Convolutional Networks and Faster RCNN
SU Hong-chao, Ying Hu, Zhu Guoqing, Chuyue Zhang · 2021
With the rapid development of deep learning in the filed of security, the application of deep learning in the factory to identify the dangerous behaviors of employees can further reduce the safety accidents caused by dangerous behaviors. Considering that the existing behavior recognition network cannot solve the problems of multi-target behavior recognition, slow detection speed and low accuracy. In this paper, a multi-target recognition model based on improved Two-Stream Convolutional Networks and Faster RCNN is proposed. The main framework of this model is based on Two-Stream Convolutional Networks. The optical flow in the temporal stream is replaced by the motion history image to reduce the computing cost. The LRCN network is introduced to replace the spatial stream network to enrich the spatial stream motion information and improve the accuracy. In this paper, the experimental research on the public data set KTH shows that the algorithm can adapt to the behavior recognition of multiple targets, with higher speed and accuracy.