机器视觉维修系统英文文献和中文翻译(9)

菜单

width. Thus, the technique can improve the

performance of the detection framework, especially when the quantity of figure pixels is large, but at the expense of lower segmentation resolution.

2.5.2. Segmentation domains. Another possible optimization involves the use of segmentation domains. A domain is an inpidual, rectangular portion of an image within which the segmentation algorithm is applied. Instead of using a single domain that covers the entire image, time can be saved by the appropriate use of several, smaller domains. The segments obtained from each domain are then combined into a resultant object set.

In our approach, there are two types of domains, spontaneous and continuous. Spontaneous domains are rectangular areas specified by a person as part of configuring the detection system to a particular location and setting. These domains are placed where the person anticipates that an object will appear for the first time. For example, detection of objects carried along a conveyor belt may work best if the spontaneous domains are placed where the belt’s projection reaches the image edges. The spontaneous domains remain static throughout the system execution, and segmentation is constantly being performed in those areas.

Although objects in our conveyor belt example originally appear only at the edges of the image, we still need to be able to detect them as they move throughout other parts of the image. Therefore, continuous domains are used to follow objects’ motion at times when they move away from areas indicated when the user selected spontaneous domains. Continuous domains are automatically regenerated after each iteration of detection. A continuous domain is formed as a rectangle centered around the locations of a new figure segment, with slightly more pixels in each dimension. This extra domain area around the sides of segments allows the system to continue to detect objects of interest as they move.

Since spontaneous and continuous domains may intersect, the set of domains is temporarily merged by a method that is similar to figure segment merging, as described in Section 2.4. Because the merging only applies for a single iteration of detection, this process results in a behavior of repeated splitting and merging that reflects the state of objects in the image, as illustrated in Fig. 4. It does so in a way that accomplishes two goals:

• Objects of interest are always contained within a domain (as long as objects originate in the spontaneous regions).

The domains also provide for robust tracking of multiple objects of interest. Even if two objects appear in the same spontaneous domain, they can follow completely different paths through the environment and can still be tracked, since separate continuous domains will be assigned to each one.

3. APPLICATION OF FIGURE SEGMENTS

3. l . Object classification

A robot vision system is expected to face a variety of objects of interest. In addition, the system may be expected to function differently, based upon the type of object that is present. Therefore, we have extended the abilities of our system to include a mechanism for the classification of detected objects. An obvious application of this ability is in an assembly line where a robot scans a workspace through which several different components pass. In this application, the robot would be given the task of servoing toward only a subset of the parts that it encounters.

We attempt to solve this problem by maintaining a model of the object type. This model consists of reasonable upper and lower bounds for each of the statistics gathered from figure image segmentation. In order to obtain these statistical bounds, training experiments are conducted with each object of interest in their natural setting (in several possible orientations). The orientations are selected so as to produce both minimal and maximal values for each one of the object’s statistics. The hope is that a resulting object model will have statistical bounds that are wide enough to encompass any general object motion, yet narrow enough to identify that object uniquely.