Fig. 27 Crossed-eye display of near-field (z, y, z) points.
Unlike Figure 20, no linking is being used.
4: Discussion
The following characteristics of this approach should be noted:
Spatial and temporal data are treated in a unified manner;
The acquisition and tracking steps of the conventional motion analysis paradigm are
merged into one step;
The approach is feature-based, but is not restricted to point features — linear features
that are perpendicular to the direction of motion can also be used;
The inherent positional structure evident in an EPI makes it is easier to analyze, and
hence easier to interpret;
Occlusion is manifested in an EPI in a way that increases the chance of detection because
the features are viewed over time against a variety of backgrounds;
EPIs facilitate the segmentation of a scene into opaque objects occurring at different
depths because they encode a homogeneous slice of the object over time;
There are some obvious ways to make the analysis incremental in time, and partitionable
in V (epipolar planes), for high speed performance.
With these benefits, the inherent limitations and current restrictions must be borne in mind:
Motion must be in a straight line and (as of this writing) the camera must be at right
angles to the direction of motion;
Frame rate must be high enough to limit the frame-to-frame changes to a pixel or so
(more specifically, such that the projective width of a surface is greater than its motion);
Independently moving objects will either not be detected, or will be detected inaccurately
(see [Marimont 1986a| for a discussion of the analysis of moving objects in the context of
this approach).
We are currently investigating the following areas:
Identifying and interpreting spatial and temporal phenomena such as occlusions, shadows,
mirrors, and highlights.
Characterizing the appearance of curved surfaces in EPIs.
Providing more dense depth information by, for example, tracking intensity levels.
Implementing the analysis of EPIs derived from motions where not only is the viewing
direction not orthogonal to the camera path, but the camera may be panning, following
some feature.
Making the analysis incremental in T' (time), rather than in V (by EPIs) - that is, pro-
cessing spatial images over time, as they are acquired.
Extending our analysis of connectivity between adjacent EPIs - this is best handled by not
losing the information in the first place, that is, by making explicit the feature connectivity
in space as well as in time. We are well into our work on this, having developed a 3-D
surface detection-linking process (a generalization of the 2-D zero-crossing detector-linker
employed for individial EPI analysis above) which provides us with this information. It
also operates incrementally over time.
In the longer term, it may be possible to remove the restriction to linear camera motion,
although this will mean losing the epipolar constraint.
128