Facial expression accounts for greater percentage of meanings in human interactions. Additionally, it conveniently and non-intrusively allows humans to conveytheir emotional state or social signs. Accurate recognition of facial expressions should therefore usher ways to the much dreamt human-computer interaction and smart environment. In such scenarios, computers are expected to communicate with human through a seamless and non-intrusive manner. Most researches on this subject were conducted on two dimensional imaging paradigms and have recorded a remarkable performance. However, changes in illumination and pose variations are two issues that impede the performance of such system and constrained them to a very tight acquisition condition. Three dimensional methodon the other hand is invariant to both illumination and pose variations and has additional depth information associated with it. This thesis investigates a novelapproach to expression recognition using the 3D method. A new Multi-ethnics 3Dbased facial expression database called (UPM-3DFE) is developed, which specifically addressed the issues of database ethnic distribution and subject outfit. Additionally,a novel method for automatic face detection and segmentation is also proposed. In this method, three salient points from each face image are robustly and automatically detected using face’s surface curvature map. The detected points are then used in selecting the appropriate sphere radius to segment the face. In the face alignment step, a new method is also proposed, that alignedthe face images intrinsic coordinate system to the world coordinate system. The feature extraction was accomplished using both geometrical and appearance features;distances, angles and line directions are used as the geometrical features,while local binary pattern filter was used in extracting the appearance features. In the final step, Support Vector Machine is employed to classify the selectedfeatures into their appropriate groups: neutral, happy, sad, angry, fear, disgust and surprise. The system achieved average classification accuracy of 92.1% for the line direction features, 89.9% for the angle features, 86.5% for the distance features and 76.3% for the local binary pattern features. This system competes favourably with several existing approaches compared with, and the results obtained are promising.