基于在线学习机制的人脸持续跟踪方法

Face continuous tracking method based on online learning mechanism

针对目标人脸短暂离开画面后不能继续跟踪的问题,提出将增量分层判别回归方法(IHDR)与改进的核相关滤波(KCF)跟踪算法相结合,以解决人脸持续跟踪问题。首先,提取人脸光照不变特征,增量构建人脸特征IHDR树。然后,检索IHDR树识别目标人脸,通过循环矩阵获取人脸正负样本,训练岭回归分类器对人脸进行跟踪。在人脸短暂离开画面时,重新识别目标人脸,使用识别结果重新初始化跟踪器,实现对人脸的持续跟踪。此外,针对KCF跟踪器的跟踪框尺度不能自适应的问题,对KCF跟踪器进行了改进,设置3个尺度不同的模板区域并计算响应,以响应最大的区域的尺度为跟踪框的尺度。最后,在不同光照下进行了动态人脸识别实验,识别率达到97.84%。与传统跟踪算法进行对比,所提方法能够在尺度上自适应跟踪目标人脸,并满足实时性要求。在人脸短暂离开画面的视频中,该方法亦能实现对目标持续跟踪。

Aiming at the problem that the target face can not be continuously tracked after leaving the screen for a short time, the incremental hierarchical discriminant regression method (IHDR) is combined with the improved kernelized correlation filters (KCF) tracking algorithm to solve the face continuous tracking problem. Firstly, the face illumination invariant features is extracted, and the face feature IHDR tree is incrementally constructed. Then, IHDR tree is retrieved to identify the target face, the positive and negative samples of the face are obtained through the cyclic matrix, and the face regression is tracked by the training ridge regression classifier. When the face briefly leaves the screen, the target face is re-identified, and the tracker is re-initialized using the recognition result to achieve continuous tracking of the face. In addition, for the problem that the tracking frame size of the KCF tracker is not adaptive, the KCF tracker is improved. Three template areas with different scales are set and the responses are calculated, and the scale of the area with the largest result is the scale of the tracking frame. Finally, the dynamic face recognition experiment is carried out under different illuminations, and the recognition rate reaches 97.84%. Compared with the traditional tracking algorithm, the proposed method can adaptively track the target face on the scale and meet the real-time requirements. In the video where the face briefly leaves the screen, the method can also achieve continuous tracking of the target.