To resolve the problem of large computational cost and bad real-time performance of current binocular vision stereo matching algorithm, a fast stereo matching algorithm of parallel system is presented. Firstly, based on the difference oftwo images, the images are divided into feature points and non-feature points. Then the feature points are matched using winner-take-all (WTA) method and the non-feature points are validated simply. Finally a dense disparity map is yielded. With the fact that the disparity map consists of piecewise-smooth surfaces, the computational cost is decreased greatly. Experimental results show that the extracted feature points are concentrated in the depth discontinuity regions. Compared with other area-base matching algorithms, the computational speed of the presented algorithm has been improved for one order of magnitude while the percentage of bad matching pixels is equivalent to other algorithms. Besides, the edge feature is better, and thus this method is an effective and feasible stereo matching algorithm with good real-time performance.