论文部分内容阅读
针对有限数目电机以及手工设置面部表情控制参数的局限,结合基于Kinect的主动外观模型,提出一种基于径向基函数神经网络的人形机器人在线面部表情模仿算法.在离线面部表情学习阶段,基于径向基函数网络建立前向机械模型以反映电机控制值与表情形变特征的映射关系,并进一步构建逆向预测模型以规整电机连续运动的平滑度;在在线面部表情模仿阶段,基于前向机械模型和逆向预测模型寻找最优电机值以实现机器人与表演者形变偏差的最小化,并引入权重因子调节表情模仿的瞬时相似度和电机连续运动的平滑度.最后,从均值统计和预测偏差角度验证两模型的合理性和泛化能力,并进一步讨论了权重因子对时空相似性和平滑度的影响.实验结果表明:前向机械模型形变预测偏差不超过1%,逆向预测模型电机控制偏差不超过1.5%.与Jaeckel、Trovato、Magtanong三种方法相比,本文算法在单帧表情模仿相似度以及多帧表情动作平滑度方面均具较好优势.
In view of the limitations of a limited number of motors and the manual setting of facial expression control parameters, an active face model based on Kinect is proposed to simulate human facial robot online facial expression based on radial basis function neural network. In the offline facial expression learning stage, The forward mechanical model is established to reflect the mapping relation between the motor control value and the facial expression deformation feature and the inverse prediction model is further constructed to regularize the smoothness of the motor continuous motion. In the online facial expression imitation phase, based on the forward mechanical model and Reverse prediction model to find the optimal motor value in order to minimize the robot and performer deformation deviation, and the introduction of weight factors to adjust the instantaneous similarity of expression imitation and motor smoothness of continuous motion.Finally, from the mean statistical and prediction bias of two The rationality and generalization ability of the model and the influence of the weighting factors on the similarity and smoothness of space-time are further discussed.The experimental results show that the forecast deviation of the deformation of the forward mechanical model does not exceed 1% and the deviation of the motor control of the backward prediction model does not exceed 1.5 %. And Jaeckel, Trovato, Magtanong three ways Ratio, the algorithm in a single frame similarity imitate facial expressions and face action smoothness with all aspects of multi-frame better advantage.