Cyber Infrastructure to Enable Computer Vision Applications at the Edge Using Automated Contextual Analysis

Principal Investigator: Yung-Hsiang Lu

Most computer vision solutions are designed to consider only pixels in images or videos. These solutions have to infer the environment about when and where the pixels are acquired. Modern cameras are often equipped positioning capabilities and can embed information about time and locations in images or videos. Such information can offer the context of the pixels. Contextual information may improve computer vision technologies in many ways. For example, many vehicles are expected during rush hours (time) in the downtown of a city (location). In contrast, few people are expected during semester breaks (time) on a university campus (location). Contextual information can be used to improve computer vision in many ways: First, the information can evaluate correctness: for example, an elephant is not expected in a city downtown. Second, such information may help to trim deep neural networks by removing impossible scenarios; for example, a traffic camera does not need the ability to recognize an elephant. Third, smaller neural networks may be deployed in edge devices that can perform computer vision without sending pixels through networks. 

 

Personnel

Other Faculty: Vipin Chaudhar (Case Western), George Thiruvathukal (Loyola University Chicago)

Students: Caleb Tung (tung3@purdue.edu)

Representative Publications

• See the World through Network Cameras (accepted by IEEE Computer for publication)

Keywords: computer vision, contextual information, machine learning