Development and Implementation of Human Face Alignment and Tracking in Video Streams

The paper presents a method that allows detection, alignment and tracking of a human face in a real time in video streams. To detect and to align face on an image a face shape regression approach is used. The developed method uses scanning window, a cascade of ensembles of regression and classification trees, and adaptive boosting. The same trees are used for classification whether the given window contains a face and for regression of a face shape. For face tracking a starting position for face search is taken from the found shape on the previous frame. Conducted analysis of the proposed method implementation gave good performance results but revealed shortcomings and directions for future work. Sensitivity of face detection is 78% and accuracy of face alignment is 95%. The implementation can track faces in real time with a speed of at least 20 frames per second.