Stereo matching networks encode 3D scene structure into binocular feature correspondences through cost volumes, which has important application in robotic localization and obstacle avoidance. However, the existing cost volume methods fail to establish comprehensive yet non-redundant correlations between binocular features, leading to low disparity accuracy. To address this issue, this research introduces epipolar geometric constraints into the cost volume computation and proposes a plug-and-play fusion cost volume and cost aggregation method by leveraging the complementary advantages of multiple cost volume types. First, the proposed fusion cost volume calculates global dot-product correlations and local grouped dot-product correlations of feature vectors within the epipolar co-projection region simultaneously, which ensures the comprehensiveness of feature correlations information and effectively avoids information redundancy. Then, during information aggregation, this research combines traditional aggregation methods with the characteristics of the fusion cost volume to propose an adaptive weighted dimensionality reduction method based on depthwise separable convolution, which addresses the dimensionality imbalance and computational efficiency issues in the feature aggregation process of the fusion cost volume. The stereo matching framework integrating the proposed method is named FusionStereo, and experimental verification is carried out on the benchmark data sets. Results show that the BAD3 of FusionStereo is 1.55% after training on KITTI 2015, and the BAD1 of Middlebury cross-domain evaluation is 17.1%, which is significantly better than other cost volume methods.