基于轻量级深度网络的动态人脸跟踪方法

Dynamic face tracking based on light-weighted deep network

为解决人脸跟踪过程中特征提取速度慢、跟踪实时性不足,尤其在人脸快速大范围移动及多人物视野下极易导致跟丢或错跟的问题,提出基于轻量级深度网络的动态跟踪方法。对人脸图像进行数据增强和人脸区域标注,搭建基于分散注意力机制的异构网络快速提取人脸信息;在此基础上引入观测框动量跟踪算法,有效捕捉包括骤移、骤停、大范围快速移动等在内的人脸信息,结合多元回归函数提升跟踪稳定性。实验结果表明,在跟踪效果上更流畅、高效,实际跟踪速度达52帧每秒,对人脸快速移动、遮挡、复杂场景变换等都有较强的鲁棒性和较高的实时性。

A dynamic tracking method based on lightweight deep network was proposed for the purpose of solving the problem of slow feature extraction and insufficient real-time tracking during face tracking,especially in the case that face moves rapidly in a large range and face moves in multi-character vision,which extremely easily leads to the issue of missing tracking or wrong trac-king.The face image was enhanced and the face region was labeled.A fast extraction of face information based on split-attention heterogeneous kernel-based residual networks was constructed.On this basis,the momentum tracking algorithm of observation frame was introduced to effectively capture the face information including sudden movement,sudden stop and large-scale rapid movement.The tracking stability was promoted through combining multiple regression functions.Experimental results indicate smoother and more efficient tracking effect,and the real tracking speed reaches 52 frames per second,which shows strong robustness and high real-time performance for fast face movement,occlusion and complex scene change.