A fast and accurate weld seam feature point recognition is the key of the weld seam tracking system based on structured light sensing． For the streak discontinuity caused by the deformation of structured light stripe line at the weld，tracking tests were carried on butt and lap welds of stainnless steel board，and an improved mean-shift algorithm was proposed to extract the feature point of a weld seam． Unlike the traditional algorithm，the improved algorithm eliminated the process of extracting fringe center line and fitting fringe lineand the feature points of welding seam are identified by drift． In order to prevent from back-shifting，the search direction of the algorithm was limited． To improve the running efficiency，a shifting accelerating factor was introduced． The test shows that the improved mean-shift algorithm can effectively recognize the feature point of a weld seam and saliently improve the accuracy and real-time performance．