Abstract This paper discusses the development and integration of a stereo vision system with an industrial robotic arm for welding operations. In order for the effects of lighting and background noise to be overcome, laser diodes are used to highlight the welding area. The distortion of the laser lines is used for the identification of the welded parts’ flange and for the correction of the robot’s path so as for the welding spot to be properly located in this area. The system is deployed on an assembly cell that produces a part of a commercial vehicle door. The performance and accuracy of the system are validated in this case. 1. Introduction Machine vision is used in many assembly and welding applications in order to provide image-based solutions for a robot control process. Applications such as quality inspection, geometry measuring, area/volume measurement and robot path correction are the ones that are undertaken by vision systems   [1].  A first classification of vision systems is the distinction between 2D and 3D systems with the first ones being the most widely used. 35708
The 2D vision systems work well for parts that can sit on a flat surface, and enables them to be picked (e.g. by a SCARA robot).  In many cases though, where the assembly can take place in a three dimensional space, it is possible that a small portion of the problem (e.g. picking or placing of the part) be treated as a 2D problem   [2]. Although 2D systems have proven very reliable, fast and accurate in past applications, the development of innovative flexible assembly technologies and the increasing product variability is continuously posing new challenges, in terms of process adaptation capabilities  [2]  [3]. Experience with such applications has proven that machine vision for robot applications requires simple-to-use and robust solutions. Usually robot control applications require 3D information on the workspace and manipulated object in order to implement pick and place activities   [4].  Especially in the automotive sector, the complex product geometry and the product variability, which are dictated by the need for continuous product re-designs, result in the need for continuous line reconfigurations in order for market demand to be met  [5]  [6]. Moreover, the use of highly automated assembly lines that are mostly robot based and the introduction of assembly technologies, which are performed in the three dimensional space (e.g. remote laser welding), require monitoring systems capable of fast and accurate, real time process control   [7].  The robotized welding has a relative high flexibility and covers a wide range of welding applications, particularly in the automotive industry    [3] [5]  [7].  Seam tracking and inspection are the most common tasks. 
However, the complex geometry of the components to be assembled, requires that these tasks are carried out in a 3D space, thus necessitating the use of vision systems with 3D capabilities  [1] [2] [8]  [9].  There exist several types of three dimensional vision systems, including active systems (also known as   structured light systems), passive systems (also known as Disparity Vision or Stereo Vision systems) and single camera vision systems.  The active vision systems make use of a camera and a laser beam source, which is used to projecting a light pattern (grid, lines, colour patterns etc.) onto the surface of the object to be measured.  This projection is captured on the acquired image and is used for calculating the point coordinates by means of triangulation.  In a similar manner, the passive vision systems use two cameras to acquire two different pictures of the object.  Then, the points of interest are matched in both images and their locations on each image are used to derive the real world coordinates with respect to the camera’s reference system. Looking at existing systems, the Focus  Robotics, offers a stereo vision system which provides real time depth perception via two 752x480 resolution cameras with a 6cm baseline.  With the use of the SAD algorithm and an 8mm focal length lens, at a range of 0.87m, an accuracy of ±0.6cm can be achieved   [10]. The Bumblebee 2 by Point Grey uses resolutions of 640x480 or 1024x768 with a 12cm baseline and the use of sub-pixel interpolation techniques allows for accuracy of 1 cm at a distance of 1m from the object   [11]. However, the intrusion of 3D systems in industrial environments, is relatively poor and there are several examples of widely known automotive companies that are currently relying on 2D systems but foresee their transition to 3D ones in a timeframe of three to four years   [2]. 2. Problem Definition A common practice for the assembly of metal parts, being for example the vehicle door components or vehicle body sides, is the use of flanges. Flanges are thin strips of metal that are created during the stamping process in order to provide an area where the welding of two parts can take place.  Usually, the flanges of each component are adjacent and parallel to each other in the final assembly of the product, allowing accessibility to the welding equipment. The flanges in the CAD model of a door and the actual frame are shown in Figure 1.     (a)                               (b) Fig. 1. Door frame and reinforcement– CAD and actual parts. 立体视觉系统的机器人英文文献和中文翻译:http://www.youerw.com/fanyi/lunwen_33781.html