Research On Dense Stereo Matching

With the growing development and improvement of computer stereo vision, the reconstruction of three-dimensional shape from two or more images is a well known and intensively investigated research problem within the Computer Vision community. A substantial amount of work has been done on stereo matching giving rise to a variety of novel approaches attempting to improve upon existing early methods and satisfy the high accuracy demand in diversified application domains such as object recognition, robotics and virtual reality.Based on the sequence images from video in this paper,it referred to the filter image noise,the comparison and analysis for the similarity between the two images ,the research strategy based on the area in local as well.The main work to be done as follows according to the above matter in the paper:(1) The impulse noise in the input image has quite greate influence to the following dense stereo matching and an improved adaptive median filter algorithm was proposed in this paper.Firstly,salt and pepper noise could be detected via the gray value 0 and 255 ,and filter the detected noise depending on whether there is signal point around its neighborhood;the algorithm reduces the false detection rate and meanwhile protect original image details during the simulation experiment;(2) The similarity between the left and right image has still impact on the dense stereo matching afterwards.The analysis and research were made in this paper based on the grey similarity function currently and the comparison was made among SSD,SAD and ZNCC ;Then ZNCC was confirmed for the proper function according to the result in this paper;(3) The matching algorithm based on area has the feature of searching pixel by pixel and then more complex calculation;while in this paper,the seeds were extracted firstly from both of the left and right image and then rank the extracted seeds according to the ZNCC score from high to low,meanwhile the highest ZNCC score was made to region-growing firstly in the scope of the calculated search window,and the point which is over the fixed threshold was put into the queue of seed;repeat the above steps until there is no seed in the queue of seed.And finally,a dense disparity map was generated .Experiment results show that the algorithm has the premise of computing speed,a higher accuracy.