I.    CLIF

            

 

v    Columbus Large Image Format

v    EO aerial imagery around OSU campus

v    Frame rate: 2 frames per second

v    Typical sensor altitude: 7000 ft (2.1 km)

v    Sensor: 2 x 3 array of 6 cameras

                 

II.   Goal

 

v    Track objects such as cars and trucks

 

 

III.  Alignment

 

                  Features: Harris Corner                           Descriptors: SIFT                           Descriptor Matching

 

RANSAC based homography fitting

 

IV.  Background Subtraction

 

 

 

v    Gradient Suppression

 

 

V.   Tracking

 

1.         Tracking Framework: Bipartite Graph Matching

v Nodes at time t and t+1

§  Edges to nodes at t+1

v Occlusion nodes

§  Edges to occlusion nodes

v Dead end tracks from p frames back

§  Reacquisition edges

 

v Solve using Hungarian Algorithm

 

2.         Scene Constraints for Initialization

v Accurate velocity estimate is needed to obtain proper correspondences which is not available initially

v Use local contextual information to help with assignment

 

2a.       Global Velocity

v Compute all possible velocity orientations between two frames.

v Obtain histogram.

v Select mean of histogram mode as the global velocity.

 

v Compute weight as:

 

 

 

 

2b.       Neighbor Context:

            For every object to match:

v Compute vectors from it’s position to neighboring objects

v Obtain 2D histogram of orientation and magnitude

 

v Compute weight as:

 

 

 

 

3.         Graph Weights:

 

                       

4.         Grid Cells and Object Handover

            Images divided into cells (for speed and constraints):

v Cells have overlap

v Bipartite matching done for each cell

v Tracks crossing cells are linked together

VI.  Multiple Cameras

 

v Different Camera Response Functions (CRFs):

 

v Inter-camera equalization using Gamma Function:

 

v Alpha-blending:

 

 

VII. Ground Truth

 

 

            [Download sequences information & ground truth here]

 

VIII. Qualitative Results

v Comparison against nearest-neighbor greedy bipartite (without constraints):

 

 

                               Track Completeness Factor                                     Track Fragmentation  

 

IX.  Qualitative Results

 

 

 

            [Download video for Sequence 1 here]

X.   Publications

 

Vladimir Reilly, Haroon Idrees, Mubarak Shah,

“Detection and Tracking of Large Number of Vehicles in Wide Area Surveillance”,

European Conference on Computer Vision, 2010.

 

[Download paper here][High Resolution]