Abstract: Aiming at the problem of “cheating” unlocking through photos or videos in the face recognition system, a live detection method is proposed to judge whether the detected object is a real person through random instructions. In this paper, HOG feature and random forest algorithm are used to detect the face of the image collected by the camera, and the 68 feature points of the face are predicted. The 68 feature point coordinates are normalized, and in combination with the face position, the resulting relative position coordinates are used as facial gesture features. The posture features and SVM classifier combine to train the head posture estimation classifier with good classification effect. Through several experiments, the feature extraction method is optimized to further improve the detection accuracy. Finally, a random interactive command sequence is used to instruct the person to be tested to perform live detection. This method has higher robustness to head pose estimation and can effectively prevent photo and video authentication attacks.

Key words: face detection, face alignment, living body detection, support vector machine, interactive random action

摘要： 针对人脸识别系统中出现的通过照片或视频“欺诈”解锁问题，提出一种活体检测方法，通过随机指令判断被检测对象是否为真人。利用HOG特征和随机森林算法对摄像头采集的图像进行人脸检测，预测脸部68个特征点位置，把68个特征点位置和脸部位置的相对位置归一化后作为姿态的特征，提取的姿态特征与SVM分类器相结合，训练出分类效果较好的头部姿态估计分类器。通过多次实验，对特征提取方法进行优化，进一步提高检测准确率。最后，使用随机互动式命令序列对被检测人发出指令，实现活体检测。该方法对头部姿态估计具有较高的鲁棒性，并且可以有效地防范照片和视频认证攻击。

关键词: 人脸检测, 人脸对齐, 活体检测, 支持向量机, 交互式随机动作