Geoff Taylor, PhD Thesis, Monash University
|
|
Chapter 6
Multimodal 3D Model-Based Object Tracking
Object tracking is complicated by lighting variations, background clutter and occlusions that are likely to degrade performance in domestic scenes. Furthermore, arbitrary objects may contain too few or too many cues for robust matching. Conventional model-based tracking algorithms are typically based on a single class of visual cues, such as intensity edges or texture, and fail when visual conditions eventually become unsuitable. This thesis overcomes the problem of unpredictable visual conditions by simultaneously tracking and fusing multimodal visual cues (colour, edges and texture). The robustness of multimodal fusion arises from the likelihood that at least some cues will work in any given scene. For example, a low contrasting background may cause edge detection to fail, but the tracking filter can still rely on colour and texture.
Comparison of Multimodal, Edge Only and Texture Only Tracking (Sections 6.5.1-6.5.3)
The performance of the multimodal tracker (fusing edge, texture and colour cues)
was compared with single cue trackers
using edge only and texture only information in a number of real tracking scenarios.
The tracked models
were created autonomously using the proposed
light stripe scanner and object classification
and modelling methods.
To facilitate a fair comparison, the trackers were applied off-line to
identical video sequences. Three MPEG-1 videos are provided below for each sequence
to show the performance of each tracker.
The trackers were actaully applied to stereo data, and the results only show the
view through the right camera.
Tracked edge cues are indicated in yellow, texture cues in green,
colour centroid in blue, and the tracked object is indicated by a white wireframe overlay.
While the off-line results are displayed at PAL framerate (25 Hz), the actual
processing rate is about 14 frames per second.
| ||
 |
||
| ||
| ||
 |
||
| ||
| ||
 |
||