Realtime Object Tracking based on
Dynamic
Feature Grouping with Background Subtraction
[Click
here
for the CVPR'08 poster
presentation (PDF)]
Previous Vehicle Tracking Approaches
 |
|
Background subtraction
Difficult to handle shadow
and occlusion
Performance is degraded
by stationary obstacles
(traffic congestion or
vehicles stopping at an
intersection)
Fails on sudden
illumination changes
|
|
 |
|
Feature Tracking and Grouping
Detects & tracks corners and groups them
based on proximity & motion (mostly
off-line)
Corner tracking is not that reliable, for
example, in intersection video images
Are proximity & motion sufficient?
|
Limitations on computation time, detection
rates, and perspective changes
Difficult to apply to intersection video clips
Proposed Approach
Combination of background subtraction and feature
tracking & grouping
Robust background subtraction with additional
feature
cue
Better feature detection & grouping by
adding
a background
subtraction cue
Dynamic Multi-level Feature Grouping
Grouping is done on-the-fly for realtime
applications
High-quality trajectories are obtained from
fragmented
feature
tracks
Various sizes of objects are handled at the same
time
|
|
  |
Result
[Original Video]
[Result Video]
Pedestrian Detection
[Pedestrian
detection result (from the same video clip)]
 
Bicycle Detection and Classifiation
Support vector machine was applied on texture features
for classfication
[Bicycle
classification video clip]
Interactive Trajectory Extraction
User-assisted system for post-processing: 100%
detection and accurate trajectories with a small number of mouse clicks
Example #1: [raw video
clip] [trjectory
data (first
view)] [trjectory
data
(second view)]
Example #2: [raw video
clip] [trjectory
data (first view)] [trjectory
data
(second view)]
[Data format]
Publication
S. E. Shladover, Z. Kim, M. Cao, A. Sharafsaleh, and J. -Q. Li,
"Bicyclist Intersection Crossing Times: Quantitative Measurements for Selecting Signal Timing," Proc. Transportation Research Board, 2009. [PDF]
Z. Kim, "Real Time Object Tracking
based on Dynamic Feature Grouping with
Background Subtraction", Proc.
IEEE CVPR, 2008. [PDF file]
[Home]
|
