Real-time three-dimensional localization and tracking of objects in bridge erector construction environment based on multiview integration

Traditional monitoring systems face limitations due to occlusions and field-of-view restrictions, which diminish the effectiveness of safety warnings. This study proposes a framework that integrates multi-view semantic segmentation, three-dimensional spatial localization, and dynamic trajectory prediction to address the challenge of real-time three-dimensional localization and tracking of objects in bridge construction environments, thereby overcoming the limitations of conventional detection methods. The methodology employs the Segment Anything Model (SAM) for object segmentation, establishes cross-perspective associations through epipolar geometric constraints, calculates spatial coordinates via stereo vision triangulation, and implements an improved Kalman filtering algorithm that incorporates construction machinery kinematic constraints. Experimental validation demonstrates that the system achieves an average positioning accuracy of less than 10 centimeters within a 20-meter range, a trajectory completion rate of 91.4%, and a processing speed of 42 milliseconds per frame, meeting real-time monitoring requirements. Despite limitations in extreme lighting conditions and detection of small distant objects, this approach significantly enhances construction safety monitoring capabilities, providing a new technical reference for intelligent construction management.