Ontology-Based Object Recognition
Robotics & Automation Engineering

image of Object recognition example. In this image the person and the robot are identified within the scene

Object recognition example. In this image the person and the robot are identified within the scene

 
image of the Simple ontology representing objects (orange circles) and the relationships between objects (links between them).

Simple ontology representing objects (orange circles) and the relationships between objects (links between them). More complex ontologies make use of class hierarchies and inheritance rules. Hierarchy and inheritance allows for efficient storage and reasoning.

Computer vision is a large field that includes applications for:

  • Inspection
  • Surveillance
  • Recognition
  • Image processing

The Automation Engineering Section at Southwest Research Institute (SwRI) has more than 20 years of experience in computer vision.

In the past, applications have been simplistic, operating in highly constrained situations. As client needs evolve, future applications will require operation in complex and dynamic environments. It is with this understanding that we have been pursuing cutting-edge research in object recognition.

Object Recognition

Object recognition is the process by which objects within an image are identified. The applications for this technology range from autonomous robotics to automatic image storage and retrieval. Although object recognition has been an active area of research, classical approaches have failed to perform as well as human object recognition. One reason for this shortcoming is that classical methods ignore information that psychologists have shown to be critical for human recognition: context. In other words, classical approaches to object recognition classify objects one by one based on the object appearance alone, whereas humans tend to classify an object as part of a larger scene.

Objection Recognition Application

The application area chosen for this research was object recognition in overhead imagery. Overhead imagery represents a very complex recognition problem because of the wide variety of objects that they contain. Today's high-resolution images are very detailed and allow recognition of large objects such as a building to small objects such as highway signs.

Currently, people analyze most overhead imagery. This analysis relies on object appearance for object recognition as well as the object's context, which often will indicate a particular object classification. Context is especially helpful when an object's appearance is degraded or otherwise similar to another object.

This is a very time intensive process, which is quickly being overtaken by the shear volume of imagery being produced by manned and unmanned surveillance. It is clear that automated analysis of future imagery will play a key role.

Special Context Approach

Context can mean a great many things, but perhaps the most important contextual information is spatial context. Spatial context encompasses the geometrical relationships that exist between objects. A good example of spatial context is containment. Some objects contain other objects. The identification of one object and its containment of a second object may be useful for the classification of the unknown second object.

For example, an office may contain a bookshelf, and a bookshelf may contain a book. The context of the office and the bookshelf aids in the recognition of an otherwise indistinguishable book.

Simple ontology representing objects (orange circles) and the relationships between objects (links between them). More complex ontologies make use of class hierarchies and inheritance rules. Hierarchy and inheritance allows for efficient storage and reasoning.

The technical approach to applying spatial context required overcoming two technical hurdles:

  • First, the methods that incorporate spatial context with object recognition were developed. Several methods were developed, some based on statistics and others on heuristics.
  • Second, an appropriate mechanism was required to store and reason about the contextual knowledge. Knowledge storage and reasoning is a well-studied problem in the area of artificial intelligence; a software ontology provided the best combination of flexibility and efficiency for storing the spatial relationships between objects.

Results

Simulation screen shots.

 
image of Original high-resolution image used for recognition testing

Original high-resolution image used for recognition testing

 
image of the Classical classification methods (no context).

Classical classification methods (no context).

 
image of the Context-based classification.

Context-based classification.

Markers indicate objects for classification. Blue markers indicate unknown objects, green indicates correctly classified, and red indicates incorrectly classified.

When tested across different types of imagery with nearly 50 object types, context-based recognition showed as much as a 30% increase in recognition over classical methods. Context-based recognition also resulted in reduced rates of false positives, a common failure in object classification. The significant increase in recognition demonstrates the power of context in object recognition.

Related Terminology

Manufacturing Technologies  •  computer vision  •  image processing  •  object recognition  •  ontology  •  context  •  applied research  •  internal basic research  •  fuzzy logic  •  multi-agent systems engineering  •  3D imaging  •  automation engineering  •  autonomous robotics  •  automatic image storage  •  automated image retrieval


Benefiting government, industry and the public through innovative science and technology
Southwest Research Institute® (SwRI®), headquartered in San Antonio, Texas, is a multidisciplinary, independent, nonprofit, applied engineering and physical sciences research and development organization with 10 technical divisions.
07/13/16