| Vis ion s ystem has lo ng been a key research fie ld of robot techno logy. Thefunction of vis ion syste m on robots equa ls to that of the eyes for human being,which provides large a mount o f optica l information for body control. Vis ionsystem is a dominant sens ing s ystem for mobile robots for it is the source ofsurround ing information. The system work in two steps, first computes out thedepth value for each pixe l of the reference ima ge of the ima ge pair collected bytwo horizontal paralle l fixed cameras, second realized3D terrain reconstructionby us ing the depth va lue and the p lane coordinates of each pixe l. Robots do pathplanning according to the reconstructed3D terrain obtained by the vis ionprocessing syste m, whic h means the system indirectly instructs the motion andcontrol of the robots.Stereo matching is the core algorithm of3D reconstructio n of the binocularsystem, it is characterized by large computing amount, restriction betweenaccuracy and speed, poor robustness for vario us image collecting cond ition. Thelarge computing amo unt attributes to large ima ge informatio n memory costs,since ima ges is stored by each pixe l which contains multi-dimens ion vectorrecording color value of3channe l, pixel coordinate and intens ity va lue. Thoughit cost high storage space, it has single store pattern for each pixe l, therefore,image object is quite proper for paralle l comp uting. According to the abovereasons, a paralle l based stereo matching system us ing CUDA was proposed toaccelerate the image processing speed. The syste m was devised based on paralle lprocessing functio n of GPU hardware to realize real-time stereo matching. Thesystem contains paralle l a lgorithms for image read and write, pre-processing,disparity ca lculation and d isparity re fineme nt. In order to obtain better matchingresult, the system set several adjustable parameters to control the final outcome.Through the study of traditiona l stereo matching algorithms, the proposedmethod comb ines the advantages of the speed for local method and the accuracyfor global method. Based on paralle l storage feature of GPU, a d isparity va luecalc ulation method us ing feature points and Bayes maximum posterior distrib ution and a disparity refine method us ing revised cross-based adaptiveweight aggregation were proposed. The disparity ca lculation method firstlycalc ulate out robustly matched feature points of image pairs us ing horizonta l andvertica l Sobe l operator, then us ing these points formulate a sparse disparity sp ace,after that, a Bayes prior model was built on that space, then us ing Bayesmaximum posterior distribution so lved out the va lue for each pixe l in itscorresponding d isparity space unit. After basically calc ulate out a disparity map,bad p ixe l detection was performed for the matching result, then us ing d isparityfilling process to cover bad pixe ls. The Disparity refine ment method fills the badpixe l by convo lutio n between adaptive weight and the basic disparity va luecalc ulated above. The adaptive we ight is derived by two-step aggregation infixed window of, the first step aggregates weight on surrounding pixe ls toassistant arm shaped of#·, the second step aggregates weight on assistant armsto the cross shaped main arm. The two paralle l realized methods ensure the speedand accuracy of the matching results.To evaluate and analyze the proposed stereo matching system, a binocularvisio n experiment hardware platform was engaged to testify the performa nce ofthe syste m. Using Middlebury standard test image pairs and practica l collectedimage pairs, the proposed system was verified by high speed and accuracycompare with traditiona l stereo matching a lgorithms. As the experime nta l resultsshown, the system tota lly fulfills the real-time requireme nt for practicalimplementation. |
