Satellite Assembly Target Recognition Method Research Based On Local Binary Feature Description

Compensates nonlinear transformation such as scale,perspective,rotation.Identifies typical assembly and workpiece target accurately and effectively are basis and guarantee of satellite ground install.Local feature can keep robustness of complex nonlinear transformation influence and express significant information in image region.Binary vector uses hamming operation and digit calculation to improve recognition efficiency.Looks forward to related technology assist robot system to realize installation work.This paper is supported by the National High Technology Research and Development Program of China,National Natural Science Foundation of China.Based on achievement of traditional feature description and target recognition.From scalar quantization index,sampling model creation and optimal projection transformation.Carries out theoretical analysis about overall recognition algorithm and core step.Uses simulation and real experiment to evaluate distinguish performance,unique property,real-time characteristic.Checks up accuracy and effectiveness about theoretical analysis.In order to meet the requirement of identifying standard workpiece and typical assembly accurately and effectively.Innovative work of this paper includes:1)Scale invariant feature transform faults with high computation complexity.Puts forward typical target recognition method based on scalar quantization and inverted index.Thresholds floating-point type initial data to improve matching efficiency.Stores higher correlation data with inverted list to compress time.Selects mask information to shade approximation mid-value and reduce influence of external nonlinear transformation.Combines ratio matching criterion and random sampling to identify standard workpiece.Experiment compares distinguish and real-time performance from standard workpiece recognition,mean average precision and running time.Ensures approximate accuracy with SIFT and tests effectiveness about SQIF.2)FREAK prones to invalid matching and error recognition.Puts forward typical target recognition method based on improved fast retinal keypoint descriptor.Analyses the limitation of retinal keypoint descriptor.Divides gray and gradient direction information to obtain more description ability.Designs sampling model and point combination for more target corner and edge feature to enhance unique attribution.Inspires by nearest neighbor matching criterion and K-D tree search strategy,queries initial feature points;Based on random sampling and correct matching rate identifies standard workpiece.Experiment compares the distinguish and real-time performance from layered step and mapped model parameter,correct matching rate,standard workpiece recognition and running time.Ensures approximate efficiency with FREAK and verifies accuracy about ImFREAK.3)Description and recognition theory with floating-point type and binary type often exists real-time or distinguish performance defect,Puts forward typical target recognition algorithm based on optimal projection transformation descriptor.Solves optimal projection variable according to decomposition principle of gramm-schmidt orthogonalization.choices transformation parameter with label data by interval sampling process.Inner product operates to establish binary description vector.Through ratio matching criterion and BBF queries matching points;Based on random sampling and correct matching rate identifies typical assembly.Experiment compares distinguish and real-time performance from projection variable parameter evaluation,receiver operating characteristic curve,external assembly recognition and running time.Tests accuracy and effectiveness about OPTI.4)Considers the core position of description in recognition process.Explores description algorithm of multi-scale rectangular area fast optimization and selection.Inspires by multi-scale space basic theory.Through continuous regional segmentation and gaussian smoothing to improve scale-invariant property.Extractes candidate area through constraint condition of compact size factor,rank number and overlapping range.Calculates response value and selects subset with stronger distinction and lower correlation to constitute description vector.Experiment verifies distinguish ability and real-time characteristic about description approach.5)Considers the critical step of matching in recognition process.Further explores fast matching strategy with multi-level quantitative difference Parameter data embedded.Combines ORB detection and description basic theory.Uses similarity judgment and evaluation rule or presupposed query and search matrix to achieve initial matching result of center feature,determines matching level according to quantitative difference data include direction attribute,space scale or coordinate position.Selects subset with high credibility and reputation from initial matching result.Experiment verifies distinguish ability and real-time characteristic about matching strategy.