Biomedical Engineering Reference
In-Depth Information
quantities of raw data. Parallel processing techniques that combine image sensing and
motion computation into a single stage are becoming an increasingly attractive alternative.
A variety of motion sensors based on analog very-large-scale-integration (AVLSI) technol-
ogy have been developed for motion detection purposes.
Most reported AVLSI motion sensors incorporate silicon photoreceptors and parallel
processing circuits on a single chip (84; 85). Some problems that have been encountered
include limited dynamic range and low spatial resolution. Simple photoreceptors in bio-
logical vision systems can respond to stimuli over extremely wide ranges. In addition,
their visual pathways are highly parallel-connected, allowing simple processing elements
situated within the pathways. It is difficult for silicon-based AVLSI systems to achieve
such high spatial resolution within a size comparable to a biological system. Furthermore,
the use of a planar imaging surface and perspective projection optics often leads to unde-
sirable image distortion or bulky vision systems. Instead, using curved or spherically
shaped retinas can provide uniform sampling across the entire field of view. For these rea-
sons, biological materials have received much attention from artificial vision researchers
because they can perform many complex functions at the molecular level.
Experimental results reveal that the photoelectric signals generated by bR photorecep-
tors exhibit differential photosensitivity. Both wavelength and intensity of incident light
influence the signal amplitude. The photoresponse peak is linearly proportional to the
illumination intensity over a wide range. These unique photoelectric properties make bR
a viable material for real-time image-processing applications. Most proposed bR-based
devices that exploit these photoelectric properties are simply light detectors or image sen-
sors that have not been efficiently incorporated into vision systems (49-51). Image data
acquired from biological vision systems is highly compressed to eliminate unnecessary
information, thereby improving transmission and processing speed. In this section, a
motion detection sensor that uses a bR photoreceptor array is described as a sample
application. It is capable of detecting both speed and direction of a moving light source.
17.5.2
Motion Detection Algorithm
Algorithms for motion detection can be divided into two major categories: intensity-based
algorithms and feature-based algorithms (51). Intensity-based methods use image irradi-
ance directly to estimate optical flow. The two most popular types of intensity-based
motion algorithms are those based on gradients or on correlation. Feature-based methods
are commonly used in computer vision to extract unique features from an image, such as
edges, corners, or patterns. Motion is estimated by tracking these features over time. Most
motion detection algorithms tend to be computationally expensive and difficult to realize
in hardware.In contrast, motion sensors based on biologically inspired vision models pro-
vide a simpler architecture and are more suitable for hardware implementations.
The Reichardt motion model is a model to describe an insect's visual response; it is one
of the most common motion detection models implemented in hardware (86). As shown
in Figure 17.20, this model employs a bilocal delayed-coincidence detector that compares
light intensity at two locations separated by a time delay. It can resolve both direction and
speed of the observed motion. The signal from one photoreceptor is correlated with the
delayed signal from its adjacent photoreceptor. Spatiotemporal filters extract spatial and
temporal edges from the input signals, while removing the DC components. When the
delayed and nondelayed signals arrive simultaneously at the multiplier, the resulting sig-
nal indicates a particular direction of motion. Subtraction between the two half-detector
signals results in an output, which is positive for rightward movement, negative for left-
ward movement, and zero for no motion or flickering stimuli. The delayed time between
adjacent photoreceptors contains the movement velocity information.
