Single Face Scalable Fast Face Detection Algorithm

Face detection is a key to face tracking and recognition. Currently, a research emphasis is to improve the detetion efficiency and detection accuary. With the main application of high-revolution remoter in video capture, AdaBoost algorithm is the main face detection algorithm, whose fundamental problem is the large amount of detection windows placed in the cascade classifier.This paper proposes a MFHE algorithm. Started from the global target motion and local face feature, according to the correlation between the background and the current frame to extract the target motion area using the patch as the calculating unit; combining with rigid motion based on the motion features to obtain the set of candidate windows. To improve the efficiency of the algorithm by reducing the amount of detection windows placed in the cascade classifier. Simultaneously, considering in the practical application scenarios, cameras or any sensors exist the minimum or maximum identified objects. If the scope is beyonded, the detection process will be insignificance. Based on the MFHE algorithm, we propose to calibrate the camera interior and exterior orientation parameters by Direct Linear Transformation (DLT) with a single camera, establishing a monocular ranging modle to measure the horizontal distance between the object and the camera. To effectively reduce the windows traverse range, according to the pinhole imaging principle to calculate the minimum and maximum identifyible window size to get the window traversal interval, and train the cascade classification with the lower limit. Finally, to optimize MFHE algorithm, the minimum and maximum window size are embedded in the MFHE algorithm to get the window set of candidate face. The window in the set is palced in the cascade classification one by one to achieve a rapid face detection.In the compared experiments, the results of V & J algorithm are treated as a criterion,our alogirthm is superior to the V & J algorithm, S & A algorithm and variation feature algorithm in the detection speed. On the video frame of 640×480 and 1280×720, the average detection speed is 28fps and 6 fps respectively which is 1/4 of V & J algorithm and 1/3 of S & A algorithm and the accuracy is slightly lower than V & J algorithm, and higher S & A algorithm. Consindering the accuracy, our algorithm is adapt to some place where the illumination changes in scenes, including the airport, the entrance of sub-way or supermarket.