Research On Depth Perception Method Based On Binocular Vision

How to enable computers to have depth perception capabilities like humans has always been one of the core issues in the field of computer vision.Binocular vision,as an important form of image depth perception,is widely used in robot navigation,advanced driving assistance systems,industrial inspection,face recognition,cultural relic protection,and other fields due to its low cost,easy implementation,and wide applicability.Therefore,binocular depth perception(hereinafter referred to as binocular depth perception)based on binocular vision has become a research hotspot in the current academic and industrial circles.At present,binocular depth perception is mainly divided into two types: passive and active methods.Passive binocular depth perception uses natural texture information on object surfaces to find corresponding pixels in two images through stereo matching and calculate disparity to obtain depth information.Active binocular depth perception adds structured light projection on the basis of passive binocular depth perception to enrich the matching features of weakly textured(or textureless)surfaces,thus meeting the application requirements of high-precision 3D object reconstruction at close range in indoor environments.In order to improve the accuracy of passive binocular depth perception in weakly textured regions,depth discontinuous regions,repetitive texture regions,color-similar and large-disparity continuous variation regions,and the robustness to lighting or exposure changes,and to improve the accuracy and efficiency of active binocular depth perception based on random speckle projection,this paper conducts in-depth research on binocular depth perception methods.The main work and innovations are as follows:(1)In order to improve the accuracy of passive binocular depth perception in weakly textured regions and depth discontinuous regions,this paper proposes a stereo matching algorithm based on guidance image and adaptive cross support region.Firstly,in order to obtain more accurate matching cost,a matching cost calculation method based on improved image gradient information is proposed.This method combines the gradient information of the rectified input image and the corresponding guidance image,calculates the similarity between pixels from the horizontal and vertical directions respectively,and constructs the matching cost function by weighting fusion with two classic methods,AD(Absolute Differences)and Census transform.Secondly,in order to comprehensively and effectively correct the outliers in the disparity map,a multi-step disparity refinement method based on adaptive cross support region is proposed.Experimental results show that compared with the local stereo matching algorithms and deep learning-based stereo matching algorithms submitted to the Middlebury evaluation platform in recent years,the algorithm proposed in this paper has higher matching accuracy in weakly textured regions and depth discontinuous regions,and better overall performance in disparity accuracy.(2)This paper studies the problem that the accuracy of the passive binocular depth perception method based on ACGF(Adaptive Cross Support Region Based Guided Filter)is not ideal in the region of color-similar and large-disparity continuous variation.Firstly,the reasons for the problem are analyzed,and then based on the construction process and structural characteristics of the adaptive cross support region,this paper introduces orthogonal weight for each pixel in the support region relative to the central pixel,and proposes a cost aggregation method ACGF-OW(ACGF with Orthogonal Weight).In addition,in order to improve the computational efficiency of ACGF-OW,a fast weighted aggregation computing method is proposed by utilizing the decomposable calculation feature of orthogonal weights.Experimental results show that compared with direct calculation under the same disparity result,the proposed fast weighted aggregation computing method reduces the average calculation time on the Middlebury evaluation platform training set images by72.7%;compared with ACGF,ACGF-OW can significantly improve the disparity accuracy in the region of color-similar and large-disparity continuous variation with less computational time growth,and effectively improve the matching accuracy in weakly textured regions,repetitive texture regions,and depth discontinuous regions.The stereo matching algorithm based on ACGF-OW has better overall performance in disparity accuracy and robustness than existing stereo matching algorithms.(3)In order to improve the accuracy and efficiency of the active binocular depth perception method based on random speckle projection,this paper proposes a spatiotemporal matching cost function with multi-information weighting fusion.The function can comprehensively utilize the spatial and temporal domain information contained in multiple pairs of speckle stereo images modulated by object surfaces to calculate the matching cost.In addition,in order to obtain the values of all parameters in the spatiotemporal matching cost function,this paper designs a parameter optimization method based on differential evolution algorithm.In this method,since there is no publicly available speckle stereo image dataset for training,this paper explores a new strategy,which uses passive stereo vision datasets with ground truth as training data,and then uses the obtained parameter values for stereo matching of image pairs under speckle pattern projection.Experimental results show that compared with the existing spatiotemporal correlation stereo matching algorithm STZNCC(Spatiotemporal Zero-mean Normalized Cross-correlation),the spatiotemporal matching cost function proposed in this paper has higher depth information acquisition accuracy and computational efficiency.Meanwhile,it shows good application prospects in the field of 3D face reconstruction.(4)This paper studies the application of the active binocular depth perception method based on random speckle projection in fast 3D face acquisition.In order to shorten the speckle image acquisition time,this paper uses a speckle image acquisition device based on three fixed speckle patterns projection.In addition,in order to further improve the efficiency of the 3D face reconstruction algorithm based on the spatiotemporal matching cost function,this paper first analyzes the parallelism of the main steps in the algorithm,and then proposes a parallel computing process and implements GPU acceleration of the algorithm.Experimental results show that under the condition of basically maintaining the same 3D data accuracy and reconstruction effect,the execution efficiency of the algorithm is improved by more than 10 times,and the calculation speed meets the real-time requirements,which can satisfy the application requirements of fast 3D face acquisition.