Dangerous Behavior Recognition Based on Pose Estimation and Action Analysis

In order to identify dangerous behaviors such as fighting in public area, proposed a deep learning model that combines human post estimation and action analysis for behavior recognition. The model takes video frame sequence as input, first make use of HRNet as backbone network to detect human body joints and generate human pose frame sequence, and then, make use of recurrent neural network for action analysis from the pose frame sequence, thus determine whether there is dangerous behavior in this video. In the first part, optimization was made on HRNet so as to reduce its parameter and improve the accuracy of joint location. In the second part, introduced Cubic LSTM as a comprehensive model for action recognition, which analysis the motion of human joints from both temporal sequence and spatial sequence, thus achieve better inference score. Experiments show that the recognition accuracy of the proposed method can reach 92.8%. In the last part, a dangerous behavior recognition system is developed based on this model, which use a monitoring host to analysis the captured video frames from cameras, and identify dangerous behavior from them automatically, thus serve for public security tasks.