At Evolution Robotics Retail, we focus our research and development efforts around improving our object recognition and computer vision capabilities. We are fortunate to have on staff some of the leading computer vision experts in the world, many of whom were part of Cal Tech's Vision Lab. Our technology development efforts revolve around improving our core vision algorithms and in seeking ways to improve their memory and computational usage.

First, is the choice of descriptors it uses to encode unique visual patterns such as the corner of an object or the print on a label. As the most distinct regions (called features) are localized in the image, unique descriptors are computed for each of them. Several hundred such features are automatically extracted and stored in a database to describe the unique patterns in each image.

Example of a recognized image with the features highlighted.

Second, is the ability to analyze a new image to collect sufficient evidence to reliably find a match within an extremely large set of possible candidates. The algorithm that ViPR uses to select the correct candidate is similar to a voting mechanism: each feature votes for the candidate which includes a similar feature (e.g., a corner feature in the new image that matches a corner feature in a trained image). The correct candidate will receive the largest number of votes since most of the features will be in agreement; however, a single or a few votes might be incorrectly cast on wrong candidates. The likelihood that a large number of votes are cast on the wrong candidate is small, demonstrating that the algorithm is very reliable in selecting the correct match.

Recognition with partial occlusion	Recognition with different position
Recognition at a distance	Recognition at an angle with partial occlusion

Third, is the ability to do all this computation in an extremely efficient manner: recognition happens in a fraction of a second when searching a database of several hundred patterns.

• 80-100% recognition rate depending on the character of the objects to recognize
• Works for a wide range of viewing angles, lens distortion, imager noise, and lighting conditions
• Works even when a large section (up to 90%) of the pattern is occluded from the view by another object
• Can simultaneously recognize multiple objects
• Can handle databases with thousands of visual patterns without a significant increase in computational requirements (the computation scales logarithmically with the number of patterns)
• Using a 1.4 GHz PC, ViPR can process 208 x 160 pixel images at approximately 14-18 frames per second

LaneHawk™ uses ViPR technology to recognize grocery items by analyzing the printed patterns on their box, instead of using the barcode.

Lighting/Camera Unit

Evolution Robotics Retail has future plans for the integration of our ViPR® technology into additional products for the global retail market.

For more information on this technology, contact Neva Garcia at sales@evoretail.com or (626) 229-3197.