This paper presents a practical application of visual odometry (VO). Visual odometry applications are computationally expensive due to the frequent and large number of required data processing. In the present work the application is implemented in a graphics processing unit card (GPU) using compute unified device architecture CUDA and OpenCV libraries, allowing real time processing with a speed of 30 frames per second. The algorithm begins with the capture and processing of stereoscopic images to find invariant interest points (keypoints) using the GPU-OpenCV speed-up robust features (SURF) library implementation. Stereoscopic image points are projected in the Euclidean space to yield 3-D estimates of the robot's translation and rotation movements. The real time VO algorithm is applied in a practical odometry estimation in a robot's outdoors navigation experiment.