Research On Object Recognition Of Mobile Robot Based On Interest Point In Image

Object recognition of robot is one of the most basic problems. Vision sensor hasmany advantages compared to other sensors, such as, stronger ability to get informance,shorter sampling period, less effects by magnetic field, lower mass, smaller size andlower cost. Human being can recognition object by extracting effective feature fromsurrounding environment within a short time, but it’s very difficult for machine vision.It’s vital for visual recognition to extract information which can represent for object byusing technologies such as image processing and image understanding. Interest pointwhich is characterized by low computation and high repeatability is such a point that islocal feature of image. Interest point is widely used in the field of object recognition.The main work in this paper is researching on the technology of object recognition ofmobile robot based on interest point in image. Main work that has already been done isas followed.First, the paper introduces briefly the methods of object recognition of mobilerobot and the state of research, and introduces the state of object detection andrecognition based on interest point in image at home and abroad. In order to realize theintelligent of robot, self-charging is one of the basic functions.When robot want tocharge battery, the first step is to find outlet. Visual technology is selected to recognizeoutlet considering existing robot projects of laboratory.Secondly, the Harris and SIFT algorithm which have been widely used areresearched in this paper. What’s more, the principles are described indetail. Theadvantage and disadvantage of two methods by applying them to the detection anddescription of outlet is analyzed. In that case, SIFT algorithm is selected as thealgorithm which is used for detecting and recognizing object because of goodrobustness. The best parameters are selected after many tests. When matching interestpoint, appropriate similarity measurement and retrieving strategy which is using forsearching similar interest point in database is decided by comparing different methods.Threshold which is using for matching is determined by statistical test. After processingthe video which is acquired off-line，the object is recognize.At last, Bidirectional-matching is proposed to improve the accuracy of imagematching which reduce the reliability of object recognition, after analyzing the principleof sift-based matching. In order to improve real-time performance, the method which isinspired by the progress of human being recognizing object, recognition process fromcoarse to fine, is proposed. The idea is realized by using different module images whichhave different size. The idea is proved feasibility by putting it into practice. Object isrecognitized on-line by using GPU which can process complicated calculation.More different objects can be recognition by using this method in the paper.