For info on how to attend this videoconference, please email khari-elijah.c.jarrett.TH@dartmouth.edu.

"Coarse Localization for Multiple Object Tracking In Video with Minimal Feature Analysis"

Abstract

Current work in object tracking typically detects entities in each individual frame, before linking the individual static detections across frames. (Further expense occurs when tracking multiple objects.) These standard approaches require the same feature analysis methods to be applied repeatedly, frame by frame, and these static detection methods are not rapidly improving.

This thesis describes a new multi-step framework, derived in part from biological vision processing, that tracks multiple entities by first performing extremely coarse localization of entities, based strictly on low-frequency motion information, and only subsequently applying more fine-grained feature analyses to specific regions of the video. Spatiotemporal carvings of video data, tubes, consist of coarse motion models for each candidate foreground entity in a video. Detailed spatial features are then examined only within selected regions of the tubes.

After a brief discussion of current video processing state of the art, this thesis:

1. introduces the new framework, including how "bottom-up" motion-fit ellipses are hierarchically
   constructed into tubes;
2. shows multiple separate "top-down" methods for incrementally improving the initially identified coarse tubes, including instances of engulfing, subtube creation, and tube merging across temporal gaps; and
3. initial demonstrations of the bottomup, top-down, and combined system.

The results indicate a highly promising approach to recasting video processing in terms of very inexpensive initial components that are judiciously combined with small, directed, expensive verifications.

Thesis Committee

• Laura Ray, PhD (Chair)
• Richard Granger, PhD
• Eugene Santos, PhD
• Sangmin Oh, PhD

Contact

For more information, contact Theresa Fuller at theresa.d.fuller@dartmouth.edu.