Min-Pooling Cost Aggregation for Semi-Global Matching of Stereo Vision Processor

Abstract

Semi-global matching (SGM) is a low-cost method suitable for hardware implementation, while it suffers from significant memory consumption. This brief presents a stereo-vision processor that leverages a min-pooling cost aggregation method for SGM. The min-pooling method addresses this issue by eliminating redundant values and employing an up-sampling technique to restore the original size without requiring clock domain crossing. As a result, this method effectively reduces memory usage by almost half, leading to a significant improvement in large-scale depth measurement. The experimental results demonstrate that the min-pooling method enhances the continuity of disparity maps, particularly in areas with less texture, by capturing more global information and reducing noise and discontinuities. Evaluations on the Middlebury and KITTI datasets show an average accuracy of 12.19% and 5.3%, respectively, indicating a more pronounced impact on the Middlebury dataset. Resource utilization analysis reveals a 1.6-fold increase in LUT usage and a 1.5-fold increase in register usage with min-pooling, while memory size effectively reduces memory usage by 41.2% compared to the method without min-pooling.