Face Tracking in Unconstrained Color Videos with the Recovery of the Location of Lost Faces

This paper proposes a framework to track a face in multi-view uncontrolled color video sequences. The method combines Gabor responses with missing observation Kalman filter to track the face. Using this approach, it is not necessary to estimate face positions even if the detection stage fails, because the missing observation Kalman filter is able to predict the face location in the next frame of the video sequence. Literature shows that tracking approach needs a face observation, returning to initial step. This work uses a preprocessing stage that actively treats the color constancy problem. This algorithm is applied directly to non-normalized RGB space, not demanding any color space transformation. Another contribution is the identification of a range for dark skin tones, not yet identified in uncontrolled color videos. Skin-tone pixel identification reduces the number of candidates to be a face region in the image and ensures that the image region is a human face. Using skin searching region the method can predict the object motion more accurately than one that performs face searching in the whole image. The proposed framework presented encouraging results for both indoor and outdoor unconstrained test videos, considering multi-view scenes containing partial occlusion and non-uniform illumination. Moreover, its capability to recover the face location not detected in a previous frame decreases the whole runtime, making it a very attractive one.