Currency Recognition And Multi-target Tracking Algorithm

With the rapid development of the global economy,and a large quantity of currency circulating in the market,there exists a growing demand for currency intelligent recognition system in business and financial institutions.Therefore,it is necessary to develope a robust and accurate algorithm.What's more,in recent years,intelligent monitoring system plays an important role for building a safe society.Multi-target tracking in complex circumstance is a key technology in intelligent system.This thesis conducts thorough research in currency recognition and Multi-target tracking.Firstly,we propose a currency recognition algorithm based on local features and their topology.Harris corners of ROI of images are detected,SURF features are extracted to describe the local pattern and the topology relations of Harris corners are exploited to depict global structure of the currency image.Template matching is used to decide the final recognition result.We have reached an accuracy of 99.99%in the test.The algorithm is proved to be robust to distortion,stretching,illumination changes of images.Secondly,we implement and improve a mutil-target tracking algorithm by hierarchical association of detection response with online learning of non-linear motion patterns and robust appearance models.At the low level,reliable track-lets are generated by linking detection responses based on conservative affinity constraints.At higher level,the association is formulated as a MAP problem and solved by the Hungarian algorithm.We exploit non-linear motion patterns learning and multiple instance learning of appearance model to get a more accurate affinity between track-lets.For the special scenes of SED,we modified the original algorithm.We use detection result of CNN head-shoulder detector and apply Harris corners to modify the speed of the target.And in the final association of track-lets,we add a motion constraint to prevent the unsafe association.The modified algorithm is 2.5 times faster than the original one while maintaining the original performance.We applied the modified algorithm in SED,and achieved a good tracking result.