ABSTRACT
The development of robust vision is still a major issue for real-time industrial and service applications. Industrial vision sensors are generally subject to considerable noise and produce a large amount of data. Therefore, a selected number of
features
are used to measure some properties of the task. A feature is any scene property that can be mapped onto the image plane such as corner position, edge length, and centroid. Feature extraction has been historically associated with pattern recognition and refers to the process of mapping to reduce the dimensionality of the patterns. Feature extraction also improves the generalization ability and computational requirements of pattern classification. Therefore, feature extraction has received significant attention during the last two decades. However, the focus of past approaches was on off-line pattern recognition, and little effort has been spent on real-time applications of feature extraction. Considerable changes of spatial-temporal conditions in real-time applications make the task of robust feature extraction quite challenging. The main requirements of robust visual measurements, i.e., speed, accuracy, and reliability, depend on the image processing and, specifically, the feature extraction method used. To achieve a robust and effective feature-extraction process, the methods used could be tailored for the application under study. In this chapter, we will introduce some generalities on realtime feature extraction but the focus of the chapter will be on
visual servoing
application. Sophisticated techniques exist to properly process image data and to remove noise. However, these
techniques are often computationally too expensive to meet real-time calculation requirements. For example, a typical visual servoing system must operate with the sample rates of 50 to 100 Hz, indicating a calculation rate of less than 10 to 20 ms for image processing [Wilson et al., 2000]. Many previous approaches to visual servoing have assumed over-simplified environments for ease of extraction, e.g., by using artificial targets [Feddema et al., 1991]. In less structured environments, vision systems usually use sharp contrast markings in the image such as corners, holes, and circles. The focus and challenge of
many real-time vision applications such as visual servoing are to use simple, computationally feasible, yet robust feature-extraction techniques to retrieve the necessary information. The common feature-extraction task in visual servoing is to determine the image location of features such as holes and corners. This is due to the availability of these features in many industrial parts and because of the ease and robustness of their extraction. Window-based methods have been used in visual servoing systems to provide computational simplicity, reduced requirements for special image-processing hardware, and ease of reconfiguration for different applications. A few reviews of the feature-extraction methods based on specialized hardware using temporal and geometric constraints are also available in the literature (e.g., O. Faugeras’ book
Three-Dimensional Computer Vision,
MIT Press, 1993).
