基于改进3D卷积神经网络的代价聚合算法

Cost Agsgregation Algorithm Based on Improved 3D Convolution Neural Network

为了提高立体匹配的准确性与时效性,改善传统代价聚合算法计算复杂及精度不高的问题,论文提出了一种基于改进3D卷积神经网络的代价聚合算法。该方法通过运用3D卷积神经网络对由代价计算得到的代价空间进行聚合,使匹配沿着视差维度和空间维度聚合特征信息,并在此基础上将3D残差网络、3D密集连接网络引入代价聚合的计算中;最后使用视差回归对经过3D卷积处理得到的特征图进行视差精化,获得高精度的视差图。通过在标准数据集KITTI上的测试实验证明了该方法具有较高的精度与时效性。

Aiming at improving the accuracy and timeliness of stereo matching,solving the problems of high compute complexity and low accuracy caused by existing traditional cost aggregation algorithms,a cost aggregation algorithm based on improved 3D convolution neural network is proposed here.This method aggregates the cost volume,which is calculated by cost computation,by using 3D convolution neural network to aggregate feature information along disparity dimension and spatial dimension.On this basis,3D Residual Network and 3D Dense Connected Network have been drawn into cost aggregation calculation.Then disparity regression is utilized to refine the feature map obtained by 3D convolution processing.The final high-precision disparity map has been achieved.The experimental results on the standard data set KITTI show that the method has high accuracy and timeliness.