Industries are being swept up by the tide of market innovations originated from pervasive application of Robotics and Automation (RA). Given the critical role of RA in industry, it has become relevant for RA to have human-like decision making capabilities. Such enablers intrinsically require the use of flexible and robust 3D perception and control systems. In the process of automating complex automotive sub-assemblies, 3D vision-based recognition as well as grasping of complex objects is required not only for detection and categorization but also for pose estimation and robotic pick-and-place operations. In this paper, we propose a novel 3D visual perception system for sub-assembly automation based on a structured light 3D vision system. We use a novel geometric surface primitive patch segmentation approach based on Hough transforms to obtain accurate surface normal estimations from 3D point clouds for the identification of patch primitives. The most relevant primitives for our application include planar, cylindrical, conic and spherical surface patches. We extract primitive surface patches from automotive CAD models in DXF format. The models are then decomposed to simple entities such as planar polygons, vertexes and lines. Our resulting models based on the 3D point clouds are composed only by simple planes and cylinders. Our system takes advantage of the available CAD data for both object recognition and for pose estimation. Our experimental results demonstrate that we can achieve, in only a few seconds, a highly accurate pose and object class estimation.