For Atm Shade Face Detection Algorithm Research

As Automated Teller Machines become more and more common in daily life, unattended ATMs have become target of more and more crimes. This thesis proposes a face occlusion detection algorithm for ATM to secure transactions.Existing methods of moving target detection are first investigated in this thesis, and then a rough head positioning algorithm is developed. Frame differing, canny edge extraction and distance transformation are employed to obtain the motion edge of the object. The kurtosis of the result image is subject to determining the sufficiency of motion information and edge information. The upper bounder as well as both side bounders of the head is located via horizontal and vertical projection.A head positioning algorithm based on multi-resolution active shape model (MR-ASM) is proposed. Head and shoulder shape which is artificially selected to train. Then the global shape model and local gray model are obtained. For each frame in video, key points are positioned by auxiliary positioning. The local gray model indicates the direction of movement and offset of every point. The global shape model keeps the shape similar to samples by global statistical information. By MR-ASM accurate position of head and shoulder is determined and the position of head is located using ellipse fitting method.Finally, two methods of occlusion detection based on human head localization are compared and analysed. The first method introduces a series of non-linear transform principles defined by analysing the colour distribution, to fit the skin colour in an ellipse in YCrCb colour space; the second method detects face occlusion with an Adaboost-based face components detector. Because the arbitrary of human faces in actual application, demoted front eye and mouth detector are used. Both methods divide the head into eye and mouth region for occlusion detection.40video clips are recorded to simulate scenario of ATM. The face occlusion detection algorithm chooses50frames (25fps) to provide warning information through statistics. All clips are tested and the head position can be correctly determined at96%by MR-ASM combining the initial location. Both methods of occlusion detection can provide the statistically correct results.