实时抗干扰的人脸检测方法

Face detection method for real-time anti-interference

提出了一种针对实时视频流的稳定的人脸检测方法。在系统前端,结合金字塔结构、积分图和肤色覆盖率获取候选区域;级联分类器能有效检测人脸,通过改进连续型AdaBoost算法中平滑因子ε的选取方法,在加快收敛速度的同时避免过学习,合理的样本分布和弱分类器权重保证了低误检率和鲁棒性,GPU并行计算技术显著地加速了训练速度;在后端,卡尔曼滤波器用于预测和去噪,跟踪和检测结合的策略能加快处理速度。实验结果表明,该方法可以实时准确地检测人脸,能有效克服来自环境因素、姿态、表情以及遮挡的各种干扰。

A robust face detection method for real-time video stream is proposed. In the front stage of the system, pyramid structure, integral image and skin-color coverage rate are combined to obtain candidate areas; Cascade classifiers can detect faces ef- fectively, and the way to select the smoothing factor in real AdaBoost algorithm is improved to accelerate convergence and avoid overfitting, meanwhile reasonable sample distribution and weights of weak classifiers insure low false alarm rate and robustness, and parallel computing based on GPU can accelerate training remarkably; In the rear stage, Kalman filter is applied for prediction and denoising, and the combination of tracking and detection speeds up processing. The experimental results show that this method can detect faces fast and accurately, and can effectively overcome various interferences such as environmental factors, poses, expressions and occlusions.