Stereo matching, a significant research focus in computer vision, has wide-ranging applications in fields such as autonomous driving, medical imaging, and robotic navigation. To address the issue of poor matching performance in ill-posed regions within image sequences, this paper presents a stereo matching algorithm based on adaptive spatial convolution. Initially, an adaptive spatial convolution block is incorporated into the context network. Through weighted aggregation of multiple convolution kernel responses, the context network's capability to capture pathological regions in complex scenes is enhanced, achieving accurate feature representation. Attention maps of the input features are then obtained along the channel dimension. Subsequently, the algorithm uses a multiscale gated recurrent unit (GRU) network structure to optimize the initial disparity results, and the attention maps generated by adaptive spatial convolution are employed to weight the GRU's output, effectively suppressing noise and further improving the accuracy of disparity estimation. Experimental results show that the proposed algorithm achieves an average endpoint error of 0.46 pixel on the Scene Flow dataset, reducing the error by 14.81% compared to benchmark methods. On the KITTI dataset, it achieves a 3-pixel error rate of 1.40% across all regions, a 15.66% reduction compared to benchmark methods, and delivers superior disparity estimation performance. Notably, in ill-posed scenarios such as occluded or reflective regions, the algorithm effectively retains detailed image features.