A stereo matching algorithm for coal mine underground images based on threshold and weight under Census transform

Binocular image stereo matching is a key technology to realize autonomous obstacle avoidance and visual reconnaissance of unmanned auxiliary transport vehicles in coal mines. However, factors such as high dust and unstable lighting conditions in coal mines can lead to Salt-and-pepper noise in the images collected by the visual sensor, resulting in a high stereo matching error rate. Therefore, a Census stereo matching algorithm based on the combination of threshold and weight is proposed to reduce the impact of Salt-and-pepper noise on stereo matching. The main contributions include: ① threshold processing is carried out on the gray values of all pixels in the support window to remove the pixels with maximum and minimum gray values in the support window and solve the impact of outlier on the weighted fusion; ② the four diagonal pixels corresponding to the center point are weighted and fused to replace the center point pixel. Select pixel points along the four diagonal lines intersecting at the center pixel, with step sizes ranging from 1 to 3. According to the corresponding steps, weights of 0.7, 0.2, and 0.1 are assigned. Multiply the valid pixel points among these 12 points by their respective weights, then divide by the sum of the valid weights. This process yields the reference value of the center pixel point after weighted processing, addressing the issue of traditional algorithms' dependency on the center pixel of the Census transform window. Consequently, this approach enhances matching precision. The experimental results show that the average error rate calculated by the proposed algorithm is reduced by 5.64% compared to traditional Census algorithms, and reduced by 1.71% compared to the mean-based Census algorithm. What's more, the average error rate under different noise levels calculated by the proposed algorithm is reduced by 15.93% compared to the traditional Census algorithm, and reduced by 16.62% compared to the mean-based one. In non-occluded areas, the error matching rate of our algorithm is reduced by 17.19% compared to the traditional Census algorithm and 18.11% compared to the mean-based Census algorithm. The proposed Census stereo matching algorithm, which combines threshold and weight, effectively enhances the robustness against noise, reduces the error rate, and improves matching accuracy.