Résumé : Quadric surface fitting of joint surface areas is often performed to allow further processing of joint component size, location and orientation (pose), or even to determine soft tissue wrapping by collision detection and muscle moment arm evaluation. This study aimed to determine, for the femoral bone, if the position of its morphological joint centers and the shape morphology could be approximated using regression methods with satisfactory accuracy from a limited amount of palpable anatomical landmarks found on the femoral bone surface. The main aim of this paper is the description of the pipeline allowing on one hand the data collection and database storage of femoral bone characteristics, and on the other hand the determination of regression relationships from the available database. The femoral bone components analyzed in this study included the diaphysis, all joint surfaces (shape, location and orientation of the head, condyles and femoro-patellar surface) and their respective spatial relationships (e.g., cervico-diaphyseal angle, cervico-bicondylar angle, intercondylar angle, etc.). A total of 36 morphological characteristics are presented and can be estimated by regression method in in-vivo applications from the spatial location of 3 anatomical landmarks (lateral epicondyle, medial epicondyle and greater trochanter) located on the individual under investigation. The method does not require any a-priori knowledge on the functional aspect of the joint. In-vivo and in-vitro validations have been performed using data collected from medical imaging by virtual palpation and data collected directly on a volunteer using manual palpation through soft tissue. The prediction accuracy for most of the 36 femoral characteristics determined from virtual palpation was satisfactory, mean (SD) distance and orientation errors were 2.7(2.5)mm and 6.8(2.7)°, respectively. Manual palpation data allowed good accuracy for most femoral features, mean (SD) distance and orientation errors were 4.5(5.2)mm and 7.5(5.3)°, respectively. Only the in-vivo location estimation of the femoral head was worse (position error=23.2mm). In conclusion, results seem to show that the method allows in-vivo femoral joint shape prediction and could be used for further development (e.g., surface collision, muscle wrapping, muscle moment arm estimation, joint surface dimensions, etc.) in gait analysis-related applications.