Face Tracking Based On Computer Vision

Face tracking is the process to determine the correlation of frame-to-frame faces in the sequence of moving image, a key phase to process dynamic face information, which is of important value in the applications of security monitoring, video conference, AI interaction, video phone and identification. Therefore it has become an utmost hot research target in the domain of computer vision. However, it is still a big challenge for people due to the uncertainty of quantities and sizes of faces in dynamic image sequences, diversity of facial expression, variation of illumination, and complexity of the imaging environment. The single model usually cannot complete face detection under complex circumstances.To overcome the over-detection problem in skin color model for sequence images, this paper put forward a new method of face tracking based on skin color model andα–β–γfilter. It applies motion information and basic knowledge of face to detect out face on the top of body.After acquiring three frames of face positions, it useα–β–γfilter to predict and search face around prediction point. As a result, we realized face detection without the over-detection problem in the whole sequences. This method not only enhances real-time performance of face tracking, but also strengthens robustness of computation.In the face tracking process, if occlusion occurred, the search window of Camshift tracker would converge to a minimum and remain in a local area, thereby causing tracking failure. In this paper, we proposed a new tracking algorithm to overcome this problem by integrating the Camshift algorithm andα–β–γfiltering prediction. After obtaining the human face location in the first three frames by Camshift, we switch to initializingα-β-γfilter parameters, and estimate reference point of the face candidate in the next frame usingα-β-γfiltering algorithm. In the absence of occlusion, the predicted point can be used as the initial iteration point by Camshift algorithm to reduce the number of iterations and achieve real-time performance; otherwise, the value obtained byα-β-γfiltering prediction will be used to replace the Camshift output to ensure the continuity of tracking.In the face tracking process, If global searching is taken in an image, the scale of calculating of the Camshift algorithm will become larger and therefore be difficult to realize in real-time. Moreover the anti-disturbance ability of global searching will become weaker while there is a region in the background having color similar to the skin-color of face. Tracking rapidly and in real-time, a new modeling method is proposed, which combines Camshift algorithm with Kalman. The method of this paper can effectively overcome the influence imposed by the variation of illumination. Since it utilizes the Kalman prediction to minish the searching region of face and eliminate the effect caused by changing of face gesture and partial occlusion, the method of this paper has the advantages such as faster tracking speed, low calculation time and small scale of calculation, etc.It can be concluded from experimental results that the three methods have higher accuracy and reliability compared with a single skin color model or a single Camshift algorithm while tracking face so that it can meet the demand of real-time ability. Experimental results show that these proposed algorithms are effective and robust.