基于决策的人脸检测与自动跟踪方法

A Method of Face Detection and Automatic Tracking Based on Decision Making

对人脸检测与跟踪的研究现状进行了简要介绍,发现传统的MDP(Markov decision processes)跟踪算法需要手动初始化,这不利于实际场景中的灵活运用。因此,提出一种基于Viola-Jones人脸检测算法和改进的MDP自动跟踪算法。从视频序列中提取人脸的类Haar特征,采用AdaBoost算法构建强分类器,并利用级联方式将强分类器进行联合,从而提高人脸检测率。MDP跟踪算法将在线的多目标跟踪问题规划成MDP中的决策,为每一个人脸目标建立一个MDP模型,并用VJ检测器的输出来初始化该模型,将人脸的出现到消失看作是MDP模型中的状态转移,在跟踪过程中采用光流法结合Kalman运动估计提高人脸跟踪的准确性和鲁棒性,减少目标ID的分配错误。在此过程中VJ检测器作为监督指标,与跟踪器的输出进行关联。实验结果表明,该算法可以稳定地检测并跟踪场景内的人脸目标,其速率也能满足应用要求。

The present status of face detection and tracking is briefly introduced. It is found that the traditional MDP (Markov decision processes) tracking algorithm needs to be initialized manually,which is not conducive to the flexible use of the actual scene. Therefore,we propose an algorithm based on Viola-Jones algorithm and improved MDP automatic tracking algorithm. First,the Haar-like features of human faces are extracted from video sequences,and a strong classifier is constructed by using AdaBoost algorithm. The strong classifiers are combined by cascade method to improve the rate of face detection. Second,MDP tracking algorithm divides the online multi-target tracking problem into decision. An MDP model is established for each face target and initialized with the output of the VJ detector. The appearance and disappearance of human faces is regarded as state transition in MDP model. In the process of tracking,optical flow method combined with Kalman motion estimation is used to improve the accuracy and robustness of face tracking and reduce the ID switches. In this process,the VJ detector is also used as a monitoring index to correlate with the output of the tracker. Finally,experiment shows that the proposed algorithm can detect and track human faces stably in the scene,and its speed can meet the application requirements.