基于局部代价层级优化的立体匹配

Stereo Matching with Hierarchical Optimization Based on Local Cost

摘要: 随着深度学习的出现和立体匹配数据不断丰富，立体匹配结合深度学习成为研究热点。虽然结合深度学习的立体匹配算法精度上不断地获得提升，但是精度的提升带来的计算复杂度不断加深的神经网络，导致大量的计算成本，这将导致其中立体匹配的方法并不是适用于常规的CPU或者GPU运算。因此保持较高精度并降低算法计算复杂度成为当前立体匹配适用工程上的热点问题。对此本文从降低算法复杂度出发并保持算法性能的目的，提出结合深度可分离卷积的ASPP (Atrous Spatial Pyra-mid Pooling)特征提取，构建局部代价体来降低计算复杂度和内存消耗，并通过层级的方式不断地对视差进行优化，从而保证性能的稳定性。研究通过实验表明，通过局部代价的方式和深度可分离卷积能够降低算法的运行时间，以及通过层级优化和ASPP方式提取特征能保证精度水平，从而在计算成本和精度水平上有一个很好的性能表现。

Abstract: With the emergence of deep learning and the continuous enrichment of stereo matching datasets, stereo matching combined with deep learning has become a research hotspot. Although the accuracy of the stereo matching algorithm combined with deep learning has been continuously improved, the neural network with increasing computational complexity brought about by the improvement of accuracy leads to a large amount of computational cost, which will result in the stereo matching method not suitable for conventional CPU or GPU operation. Therefore, maintaining high accuracy and reducing the computational complexity of the algorithm has become a hot issue in the current stereo matching engineering. In order to reduce the complexity of the algorithm and maintain the performance of the algorithm, this paper proposes the ASPP feature extraction combined with the depthwise separable convolution, constructs a local cost volume to reduce the computational complexity and memory consumption, and continuously optimizes the disparity in a hierar-chical manner to ensure stable performance. The research shows through experiments that the running time of the algorithm can be reduced by means of local cost and depthwise separable convolution, and the level of accuracy can be guaranteed by extracting features through hierarchical optimization and ASPP, so there are many good results in terms of computational cost and precision.