基于双向GRU和注意力机制模型的人体动作预测

Human Motion Prediction Based on Bidirectional-GRU and Attention Mechanism Model

针对人体动作预测中由于受到运动速度、运动幅度等不确定因素的影响,导致预测的第1 帧动作不连续且准确预测时间较短的问题,提出一种基于双向门控循环单元(GRU)和注意力机制的端到端模型——BiAGRU-seq2seq.该模型的编码器部分采用双向GRU 结构,使数据从正反2 个方向同时输入;解码器部分采用单向GRU 结构并加入了注意力机制,使编码器输出编码成一个包含多个子集的向量序列;然后将解码器的输入和输出数据同时送入残差架构中,用来模拟人体运动速度,使预测值更接近真实值.在TensorFlow 框架下,利用目前动作捕捉数据最大的公开数据集human3.6m 进行人体动作预测实验的结果表明,文中模型不仅能极大地降低短期动作预测的误差,也能较为准确地预测出多帧动作.

Aiming at the problem that the first frame of human motion prediction is discontinuous and the accurate prediction time is short due to the influence of uncertain factors such as motion speed and ampli- tude, a sequence to sequence model (BiAGRU-seq2seq) based on bidirectional GRU and attention mecha- nism is proposed. The model encoder section uses a bidirectional GRU, which allows data to be input from two opposite directions at the same time. The decoder section uses the GRU plus attention mechanism structure to encode the encoder output into a vector sequence containing multiple subsets. The input and output of the decoder are then simultaneously sent to the residual architecture to simulate the speed of the human body and bring the predicted value closer to the true value. In the TensorFlow framework, human motion prediction experiments were performed using the public motion capture dataset human3.6m. Ex- perimental results demonstrate that the proposed model can not only greatly reduce the short-term motion prediction error but also accurately predict multiple motion frames.