Abstract. Ultrasound (US) image guidance in orthopaedic surgery is emerging as a viable non-invasive alternative to the currently dominant radiation-based modalities. Though it offers many advantages including reduced imaging costs and safer operation, the relatively low US image quality complicates data processing and visualization. We propose a novel approach for robust bone localization that integrates multiple US image features including local phase information, local signal attenuation, and bone shadowing to robustly segment bone surfaces. We demonstrate the advantages of our approach in different contexts including improved segmentation quality, increased registration accuracy, and decreased sensitivity to parameter setting. We present quantitative and qualitative validation on a bovine femur phantom and on real-life clinical pelvis US data from 18 trauma patients using computed tomography (CT) image sets as ground truth.