A parallel implementation of a proposed stereo vision algorithm for three-dimensional scene reconstruction is presented. The algorithm firstly estimates the disparity map of a scene from a pair of rectified stereo images using an adaptive template matched filter. Next, the estimated disparity is utilized to retrieve the three-dimensional information of the scene, by considering the stereo camera's intrinsic parameters. The proposed algorithm is implemented on an embedded graphics processing unit by exploiting massive parallelism for high-rate image processing. The performance of the proposed algorithm is evaluated in real-life scenes captured on a laboratory experimental platform in terms of accuracy and processing speed.