基于融合代价和优化引导滤波的立体匹配算法

Stereo Matching Algorithm Based on Fusion Cost and Optimized Guided Filtering

现有局部立体匹配算法与全局立体匹配算法相比,计算量更小、速度更快,能达到实时匹配的要求,但存在误匹配率较高、视差结果精度较低等问题。提出一种新的局部立体匹配算法,依据像素梯度信息、像素梯度的平均值及标准差计算多指标梯度代价,使算法对图像局部结构的表达更加全面。根据像素的灰度差异程度划分等级,提出7等级编码的精细化Census变换匹配代价,以有效识别图像信息之间的细微差异,弥补Census变换在相似纹理区域精度较低的不足。将多指标梯度代价和精细化Census变换匹配代价按一定权重进行融合,形成初始匹配代价,从而充分发挥两种代价的优越性。在代价聚合阶段,对引导滤波的线性系数进行自适应优化,解决固定参数引起图像局部过于平滑或平滑不足的问题,并使用优化后的引导滤波模型进行代价聚合,改善代价聚合的效果。使用胜者为王算法计算视差获得初始视差图,最后通过左右一致性检测和加权中值滤波进行视差优化,得到更为理想的视差图。在Middlebury V3立体评估平台上的实验结果表明,所提算法在非遮挡区域的加权平均误匹配率为15.7%,与Cens5、IGF、ISM等算法相比具有较高的精度。

The existing local stereo matching algorithm requires fewer computation and has a higher speed than global stereo matching,making it suitable for real-time matching.However,it has issues such as high rates of matching errors and poor accuracy of disparity results.A new local stereo matching algorithm is proposed.The multi-index gradient cost is calculated according to the pixel gradient information,the average,and standard deviation of the gradient.This allows the algorithm to better capture the local structure of the image.The proposed matching cost using 7-level coding of the refined Census transform can help identify small differences between images,which can fix the low accuracy of the Census transform in similar texture regions.Next,the multi-index gradient cost and refined Census transform matching cost are combined according to a certain weight to form the initial matching cost,thereby leveraging the advantages of the two costs.In the cost aggregation stage,the linear coefficients of the guided filtering are adaptively optimized to solve the problem of excessive or insufficient image smoothing caused by fixed parameters.The optimized guided filtering model is used for cost aggregation to improve the effect of cost aggregation.Then,the Winner-Take-All(WTA)algorithm calculates the disparity to obtain the initial disparity map.Finally,an ideal disparity map is obtained through disparity optimization through left-right consistency check and weighted median filtering.Experimental results on the Middlebury V3 stereo evaluation platform demonstrate that the proposed algorithm has a high accuracy compared with Cens5、IGF、ISM and other algorithm,with a weighted average percentage of bad matching pixels in the non-occlusion regions of 15.7%.