Abstract: During the welding process of box culvert steel bars, the structural thermal deformation caused by heat input can lead to random and nonlinear changes in the shape of the weld seam, resulting in tracking errors, which can easily accumulate and affect the welding quality and overall structural safety. Therefore, a welding seam tracking technology for box culvert steel bars under reverse compensation is proposed. Construct a Hessian matrix of the collected weld seam images using a two-dimensional Gaussian function, and extract laser stripe pixels based on their characteristic values to accurately calculate the center coordinates of the weld seam. Input the center coordinates of the weld seam into the Kalman cascade filtering algorithm, establish a state space model, recursively predict the target state using the previous estimated value and current measured value, process the coordinate data in real time, and accurately calculate the accumulated error caused by welding thermal deformation. Therefore, the center coordinate of the weld seam is transformed into a NURBS curve. The Lagrange equation is constructed by single-target point constraint algorithm to solve the weight correction of the basis function, and the trajectory is compensated inversely to minimize the cumulative error, so as to offset the influence of the cumulative error on the tracking results, so as to realize the accurate tracking of the welded seam of the box culvert steel bar. The experimental results show that the proposed method can achieve high-precision weld seam tracking while RTAMS is below 1cm. It can effectively suppress the accumulation of tracking errors, significantly improve the stability and reliability of welding automation systems, and provide reliable technical support for engineering applications.