Weighted Multi-Camera Stereo Vision Based on Geometric Relationships of Feature Edges

Improving the accuracy of visual applications for measurement has always been a hot topic in the fields of aerospace, automotive manufacturing and robotics. The multi-camera system, which can provide redundant information from multiple views to improve measurement accuracy and robustness, has been widely used. Due to factors such as pixel dispersion, image noise and mismatches, the extraction accuracy of corresponding key points in images is greatly affected. Traditional 3D point reconstruction algorithms assign the same weight to the corresponding key points in each image, which lacks consideration of the different extraction accuracy of the key points, resulting in lower measurement accuracy. In this paper, weighted multi-camera stereo vision by considering geometric relationships is proposed, which can be extended to the measurement of features with theoretical geometric relationships in projection by any number of cameras. A five-camera measurement system is designed for accuracy verification. The experimental results indicate that by reasonable weight computation, the root-mean-square errors reduce 4.46%, 5.02%, 3.38% for length, width and center distance of rectangle respectively and the root-mean-square errors reduce 3.26%, 10.69%, 2.63% for length, width and center distance of oval shape respectively compared with the least square method. Besides, the measurement robustness is also proved to have a significant improvement.