Abstract: Most previous algorithms for feature contours extraction are less efficient due to global sampling, so we proposed an algorithm based on local sampling and tracking to satisfy the real-time requirement of feature lines extraction for complex models. The proposed algorithm identifies features by multiple metrics including dihedral angle, angle defect and normal voting tensor. The credibility of feature detection can be improved by taking advantage of these metrics. The coverage of global features is achieved through local sampling and features tracking basing on initial features. During the tracking process, different types of features are identified dynamically and adaptively. The final step is to track weak feature contours from corners to ensure the integrity of the contours. We validated our algorithm through instances from the field of CAD and digital medicine. The experimental results show that our algorithm is feasible and high efficient.