Control in Robotics and Automation

314 CHAPTER 11 / AUTOMATED INTEGRATION OF MULTIPLE SENSORS

the rationalization that the results can be easily adapted to real images. Although this may

well be true theoretically, often the problems and difficulties inherent in processing real sensor

data are such that simulation-based techniques resolve few of the real problems. That is, the

distortions, sensor anomalies, noise, and so on inherent in real images so dominate the

processing problem and either the simulation-based technique is solving such a trivial

component or the anticipated mapping function from real images to simulation quality

images is at least as complex as the original IP problem. It is our view, therefore, that real

sensor data must be the focus from day one for all but function encoding verification,

graphical explanation, or preliminary concept generation. The techniques presented in this

chapter are developed and evaluated on actual multimodal sensor data consistent with

real-world robotic applications, that is, a laser range camera used for robotic navigation and

ground conductivity data used for buried waste assessment.

The following section describes the traditional approach and provides a comparative

overview of our adaptive learning approach to MSI. This approach is detailed in Section 3.

Section 4 describes the sensory target domains, including a description of the physical nature

of each sensor system (laser range camera, LRC, and ground conductivity, GC, sensor) and

their targeting environments. Section 5 presents a specific sensor anomaly common to LRCs

and our MSI approach to resolving its distortions, complete with empirical results. The sixth

section presents an empirical evalution of the performance of our automated MSI system

applied to LRC images. Section 6 validates these results empirically in the learned (LRC) and

unlearned (GC) domains. A summary section follows.

2 BACKGROUND

Although MSI has a generally accepted overall purpose, that is, to generate a more reliable

and consistent representation, the

specij'ic

goals of each MSI implementation vary from one

application to another [-4, 5, 7]. MSI techniques combine multimodal sensor images (e.g.,

sonar and visual [12], distance and reflectance [-1, 13]), multiple single modal sensor images

(e.g., multisampling, a single sensor's images displaced in time or in space [-14]), or com-

binations of these.

MSI is necessary because of basic ambiguities inherent in our current sensor imaging

technologies. Sensor ambiguities derive from two basic causes: limitations of the physical

attribute being measured by the sensor (e.g., visible light's inability to permeate opaque

surfaces, sonar edge effects) and the sensor's inaccuracies in making measurements (e.g., noise,

resolution). Ambiguity exists as long as the mapping from reality to image is not one to one.

That is, if different "realities" lead to identical images, a single image cannot reveal the

particular reality that was the truth; for example, a two-dimensional (2-D) visual image of

an opaque object cannot reveal its interior or its hidden surfaces, hence an infinite number

of "realities" would result in the same sensor image.

MSI techniques attempt to resolve some of these ambiguities by appropriately coupling

complementary images to eliminate possible inverse mappings. What constitutes the most

advantageous MSI technique is dependent on the given application domain, available

sensors, and task requirements. Multisampling is perhaps the simplest MSI technique and is

used primarily to reduce noise. Merging multiple single modal images [14] displaced in space

can improve three-dimensional information, reveal otherwise occluded areas, and reduce

orientation-induced artifacts, such as glare or sonar positional anomalities. Merging multiple

single modal images displaced in time can be used to reduce temporal effects or to isolate

them. Employing multimodal sensors permits exploitation of each sensor's strengths without