Positioning Aviation Bolts in Narrow Spaces: A Deep Learning-Based Visual Approach at Arbitrary Shooting Angles

The three-dimensional visual positioning of aviation engine rotor bolts is of great significance for automatic connection status detection. However, the internal aviation engine rotor space is narrow, and the shooting angle is limited. Existing vision-based positioning methods based on vertical shooting are insufficient to meet aviation bolt positioning requirements. In this paper, a robust bolt visual positioning method at arbitrary shooting angles is proposed. First, a deep learning-based object detection method is employed to extract bolts from both the color image and depth image. Then, an arc-support line segments (ASLS) method based on generative adversarial network preprocessing (ASLS-G) is used to detect multiple elliptical targets from the low-pixel color region of the bolt (CROB). By fusing color image and depth image information, an adaptive threshold slicing vector (ATSV) algorithm is proposed to screen out the elliptical targets corresponding to the bottom of the bolts. Finally, a symmetric depth mean (SDM) algorithm is proposed to obtain the depth value of the bolts and calculate their three-dimensional coordinates. The experimental results show that the mean absolute error of the bolt coordinate values in each direction in the bolt coordinate system is less than 0.8968 mm, and the standard deviation does not exceed 0.1636 mm, which verifies the effectiveness of the proposed method. Measurement experiments were conducted on the distance between bolts and the bolt distribution radius for a real aviation engine rotor, demonstrating that the proposed method can be applied to a broader range of real-world scenarios.