Study On Real-Time Detection Method Of Ear Video

Ear detection is the first step in ear recognition system, and the detection result will play a significant role in ear identification. Most research on ear identification is implemented on the assumption that the ear has already been detected and located. Therefore, how to detect the ear rapidly and effectively is very important.Ear detection can be implemented on static ear images and ear video sequences. Ear detection of video sequences is studied in this thesis. Presently there are a few human ear image databases, but no ear video databases are available. Since an open human ear video database is helpful to evaluate the performance of ear locating and tracking algorithm, our laboratory has established a database, which consists of 3,600 video segments collected from 120 Chinese people to evaluate the algorithm. We have collected 30 different video segments of each person under different situations --- two kinds of typical illumination change, two types of movement and three kinds of interference.To implement ear detection precisely and rapidly, we proposed a new ear detection method on the basis of some traditional methods. Firstly, ear skin probability distribution histogram was trained off-line with ear samples from an ear image database. Then an improved CamShift algorithm was used to roughly detect the ear image to get candidate ear regions. Finally, these candidate ear regions were further verified by an improved multi-step-correction AdaBoost algorithm to get precise position of ear. In this method, skin color was combined with Haar features, which can improve the speed of ear detection. There are totally 3,600 positive samples and 5,300 negative samples in off-line training, which include 12 training layers. It was implemented with C and C++ language in Windows XP operating system.Finally, this algorithm was validated by 3,600 video segments collected by our laboratory, and the results show that this method can satisfy the real-time requirement with average detection rate of 93%. This algorithm can detect ear in a variety of interference and has better adaptability to left and right ears of various gestures, shapes and sizes.