Fast Banknote Identification With RGB-D Sensor For The Visually Impaired

Banknote identification is an essential functionality of wearable assistive devices for the visually impaired.According to the literature,the traditional banknote recognition methods are mostly applied to the financial service system with regular scene,single background and fixed light source.A few of the methods involved in the field of the visually impaired,however,using RGB cameras,has poor real-time performance which cannot meet the practical application of banknote identification for the visually impaired.In our paper,an effective method is proposed to generate accurate denomination and the result will be broadcasted to testers instantly when a banknote is held in front of a wearable prototype.Firstly,close-range scenes within the length of an arm estimated 0.8m are extracted through the depth filtration with an RGB-D sensor.Potential banknote areas are determined through the classifier obtained from Haar features and Adaboost algorithm.Descriptors of these areas are matched with descriptors of standard regions prepared using speeded-up robust features(SURF),which is modified to achieve matching points with high quality.The denomination with maximum matching numbers is the recognized result and finally is reported to the visually impaired.Simultaneously,to complete an overall banknote recognition system and improve the real-time and stability of the algorithm,measures like frame differential detection,depth filling ratio judgment and multiple frame cumulation are applied to deal with the input of video stream.The approach combines RGB-D sensor,machine learning and modified SURF,making the final recognition inherit the advantage of size invariance and rotation invariance from SURF and robust with a wide variety conditions including scaling,rotation,partial occlusion,illumination change,direction reversion,motion blur,cluttered background and multiple banknotes detection.At the same time,rough banknote localization with machine learning and precise recognition with modified SURF matching,endows our method capabilities in real-time banknotes recognition for the visually impaired.