Study On Recognition Of Planar Object Randomly Placed Based On Local Feature Prediction Agreements

To recognize and locate objects that are randomly placed in a bin is a common operation in industrial automation,and it can offer the guiding information for bin picking or automatic assembly.Machine vision based objection recognition has the advantage of higher flexibility,lower cost,smaller occupation and better maintainability compared to the traditional mechanical one.In this thesis,aiming at the planar object in industrial application,a method which implements the recognition according to the prediction agreements of different local feature instances in object is proposed.First,the difficulties of planar object recognition under randomly piled condition are analyzed,and accordingly,our method,recognition based on local feature prediction agreements(LFPA),is proposed.Next,the reliability of LFPA method is inferred approximately one if there are prediction agreements achieved by no less than three local feature instances.Later,the details of LFPA method are given.Local-global prediction methods of local features in various symmetrical object and with different scale sensitivity are studied.Besides,a prediction consistency measurement is designed so that the logical relation of different local feature instances in prediction consistency can be obtained.With that logical relation,a method,fast obtaining prediction agreement sets based on the combination number of local feature index series(CNIS)is given.Then,the derivation of final recognition result from prediction agreement set containing scale sensible feature instance and not is discussed.Meanwhile,the evaluation of result priority by assigning priority weight to different local feature is proposed.Finally,through several experiments,the rules to set parameters in LFPA are illustrated and the robustness of LFPA method to illumination invariant,clutter and occlusion is also verified.