Visual odometry is a technique that estimates robot ego-motion from a sequence of stereo images.  The image below  shows an example of a real-world image and the feature matching that lies at the core of our visual odometry approach.  Extracted image features are used to establish stereo correspondences as well as  matches between two successive frames. We can then separate the stereo features based on their disparity and the vehicle speed into two sets. The set with low disparity is used to recover the camera rotation.  Based on the recovered rotation, we then use the second set of features to recover the vehicle translation.  In the image below, green flow vectors correspond to far-away scene structure (rotation only), and red to close-by structure (rotation + translation). Blue flow vectors were rejected as outliers because of noise or because of gross correspondence errors, e.g. the moving vehicle above.  This algorithm is based on work at Georgia Tech University by Dr. Frank Dellaert.  TRACLabs has just been awarded an Army contract worth $750,000 to build a compact, self-contained visual odometry sensor that will build on the work at Georgia Tech.