Giving Computers and Robots the Gift of Vision

robot-sidebar-U.S.-Navy-photo-by-Journalist-1st-Class-Jeremy-L.-Wood-e1329848117297Have you ever played an Xbox Kinect video game and felt like it was not quite reading your motions precisely? Or have you used face recognition in the Macintosh iPhoto application and had it suggest that your grandmother was you? These experiences show that, while computers’ accuracy in interpreting what they see through a camera has improved, it has not been mastered . . . at least not yet.

Technologies such as Kinect and face recognition were primarily developed in a branch of computer science called computer vision. BYU professor Tony Martinez and his PhD student Mike Gashler in the Department of Computer Science are applying their expertise in artificial intelligence to see if they can improve the accuracy of computer vision algorithms.

Gashler’s research on dimensionality reduction converts video sequences into plot points, enhancing robots’ ability to understand what surrounds them.

“Understanding what we see is an important component of human intelligence,” Gashler said. “We’re hoping to give some of that ability to machines.”

He set out to solve the problem of more accurately estimating the state of a machine. In this case, Gashler randomly changed the height and angle of a crane and then used images of the crane to estimate the height and angle at each point in time.

Because he did not have access to a real crane, Gashler used a 3-D model of a crane and a ray tracer (a virtual camera that takes 2-D pictures of 3-D models) to generate pictures of this machine. Using a sequence of 4,000 images, Gashler’s algorithm estimated the height and angle of the crane at each of the 4,000 time-steps. When compared with other methods, Gashler’s algorithm was able to estimate these values with unprecedented accuracy.

“I built a dimensionality reduction technique that is better at its purpose,” Gashler said. “The existing ones [create] a poor estimate of state; mine does it better.”

The key to the accuracy of this new approach is how it computes the changes in state. When using cameras to determine computer vision, other algorithms assume proportional distances. Gashler’s algorithm is the first not to assume objects are a certain distance; instead, it uses the points to determine the distances, and for this reason, it is significantly more accurate than older models, giving it potential to improve camera recognition technology such as Kinect video games.

Gashler has been researching dimensionality reduction for six years. This research propels Gashler through his PhD program in computer science, as he works with Professor Martinez and other PhD students and professors.

By Alysa Kleinman Posted on