Hidden Markov Model (HMM) Topologies for Visual Object Recognition

Period of Performance: 08/09/2001 - 08/08/2003


Phase 2 SBIR

Recipient Firm

IC Tech, Inc.
4295 Okemos Road, Suite 100
Okemos, MI 48864
Principal Investigator


This Phase II SBIR project will design and implement visual object recogniton modules based on Hidden Markov Models (HMM). HMM is a technique that has worked very well for speech recognition and genetic discovery. HMM's success with these problems can be attributed to its flexibility and ability to solve two problems at once, namely segmentation and recognition. This is precisely the case with visual object recognition, as well. In the HMM based object recognition technique demonstrated in Phase I portion of this project, the hypotheses formation and verification steps of traditional object recogniton architectures are merged without a mandate for a priori segmentation: HMM receives a seeet of image features in context, and in response, produces an "object word." The words may be connected to form sentences. Two or three-dimensional "object sentences" may be synthesized from object words. Hierarchies of object primitives defined in this manner further embellish the extent of the object description. Visual object recognition, especially in real dynamic environments will be of great benefit in many commercial and military applications. A specific application in the commercial domain is the audio visual speek recognition and enhancement system we will develop for automotive telematics and hand-held devices. Many additional uses grow out of this audio visual interface, such as user authentication, tracking and logging of access, and customization of user safety features, such as speed and other features of airbag deployment in vehicles. A military application to be demonstrated in Phase II is the automated analysis of video footage to label airborne objects and create their virtual representations.