Human motion prediction based on bidirectional feature sequence learning
Wei Du, Ya-Nan Yu, Qi Pan · 2022 International Conference on Computer Engineering and Artificial Intelligence (ICCEAI) · 2022
Dynamic analysis of human behavior is an important research branch of computer vision and pattern recognition field in recent years. In order to enable the computer to predict the intention of future action posture in advance by monitoring human body's early behavior, a behavior prediction model Bi-Seq2Seq is proposed in this paper, which is based on the integration of Bi-GRU unit and Seq2Seq model. Bidirectional Gated Recurrent Unit (GRU)is introduced into the encoder of traditional Seq2Seq structure to integrate forward and reverse features. It breaks away from the limitation of temporal order and effectively combines context information. The decoder part still adopts one-way GRU unit. The experimental results show that the human behavior prediction model based on bidirectional feature sequence reduces the prediction error of Smoking, Posing, Greeting and Sitting movements, with the maximum reduction up to 0.52. The short-term prediction accuracy of attitude prediction algorithm within 400ms is improved.