This paper presents a stereo vision system suitable for car assembly which requires recognition and localization of typical three dimensional industrial parts. The vision system should meet three requirements: fast processing, high accuracy and robustness. Both to reduce computational cost and to increase localization accuracy, Videorate Pipeline Sub-pixel Edge Extractor (VPSEE/vi:pi:si:/), which can extract sub-pixel accurate edges from a 512 × 512 gray image with Canny edge finder, has been developed. To recover a robust depth-map, the stereo algorithm which embodies several constraints, both to decide matching candidates and to calculate the strength of potential matches, uses information on 2D geometrical features recovered from edge segments. Experimental results show that this vision system can recognize complex industrial parts thus proving applicable for automation tasks.