基于深度信息的压缩感知人脸检测跟踪

Depth Based Face Detection and Tracking Using Compressive Sensing

传统的压缩感知跟踪是基于彩色视频图像序列中的目标跟踪,但在跟踪过程中可能会受到光照变化和旋转遮挡因素的影响,从而导致复杂环境下跟踪结果的鲁棒性不足.为了获得稳定的跟踪结果,提出了一种基于深度信息的压缩感知人脸检测跟踪算法.首先,根据改进的质心分割算法确定首帧深度图中人脸的跟踪位置.其次,根据深度信息计算出深度图中每一点对应的平均曲率并形成平均曲率图.然后,基于平均曲率图随机提取压缩特征;最后,通过压缩降维,目标邻域搜索,迭代更新特征模板,计算出平均曲率图中下一帧人脸的最优跟踪位置,实现人脸跟踪.实验结果表明,将人脸的深度信息和压缩感知特征相结合在光照变化和旋转遮挡情况下具有很好的鲁棒性,能更好的实现复杂背景下对多姿态人脸的跟踪.

The traditional target tracking algorithms based on compressive sensing uses color video sequence. It is easily affected by rotation, occlusion and illumination. Thus it has poor robustness in complex environment. In order to obtain stable tracking results, this paper proposed a novel face detection and tracking algorithm using compressive sensing based on depth information. Firstly, we identify the position of face in the initial depth map automatically using the improved centroids segmentation algorithm. Secondly, we calculate the mean curvature graph with the depth map. Thirdly, the compression features are extracted based on the mean curvature graph. Finally, by compression and dimension reduction, the feature template is updated and the best postion of face is estimated in the neighborhood space based on the mean curvature graph. The experiment results demonstrate that by combining the depth information of human face and the compressive sensing feature, the proposed algorithm has better robustness dealing with problems of rotation, illumination and occlusion.