多人脸跟踪与最佳人脸提取

针对视频人脸识别系统中同一人脸重复识别的问题,文中提出了一种多人脸跟踪与最佳人脸提取的方法。通过ViBe算法提取运动区域,缩小数据处理区域及确定执行人脸检测;利用Haar特征结合AdaBoost算法检测人脸,并根据肤色检测判断是否有误检;利用CamShift算法跟踪人脸;再使用Sobel算子得到清晰的人脸图片。实验表明,该方法下人脸误检率由2.8%降到0.2%,对于100帧视频平均处理时间从原始每帧112 ms降低到了45.6 ms,其处理速度明显提升。

Multi-face Tracking and Optimal Face Extraction

ing at the problem that repeated recognition of the same person in video face recognition system, a multi-face tracking and optimal face extraction method was proposed. Using ViBe modeling to extract the motion area, reducing the data processing area; The Haar feature was combined with the AdaBoost algorithm to detect the face, and the skin color detection was used to determine whether there was a false detection; Tracking faces with CamShift algorithm; Using the Sobel operator to got a clearer face image. Experiments showed that under this method, the face false detection rate was reduced from 2.8% to 0.2%. For 100 frames, the average processing time was reduced from 112 milliseconds per frame to 45.6 milliseconds, and the processing speed was significantly improved.