Binocular vision-based pose monitoring technique for assembly alignment of precast concrete components

Abstract

Vision-based pose monitoring for the assembly alignment of precast concrete components at long distances is often hindered by the small size of elliptical targets and cluttered construction backgrounds. To address the challenges in small ellipse identification and coordinate calculation, this study proposes a binocular vision-based pose monitoring technique for precast concrete components. First, an improved ellipse target detection algorithm is developed, integrating Omni-Dimensional Dynamic Convolution (ODConv) and a Coordinate Attention (CA) mechanism to enhance small target recognition under complex backgrounds. ODConv dynamically adjusts convolutional kernel shapes to capture multi-scale features while the CA mechanism embeds precise location information into channel attention. Additionally, The Normalized Wasserstein Distance (NWD) is employed to improve non-maximum suppression by considering both spatial distance and shape similarity, which is particularly beneficial for small and densely distributed targets. Second, an enhanced ellipse center extraction algorithm is introduced, utilizing image magnification and contrast enhancement for more accurate pixel coordinate extraction. The experiments demonstrated the superior performance of the proposed method with a precision of 97.6 %, a recall of 97.8 %, a mean average precision ([email protected]:0.95) of 70.5 %, and an inference speed of 35.8 FPS. This balance between accuracy and efficiency ensures the feasibility of real-time applications. Furthermore, field tests conducted at a hoisting site for precast concrete columns achieved positioning errors in the x and y directions within 5 mm over a monitoring range of 10 m, which confirmed the method’s practical reliability.