Tracking Vehicles in Airborne Video Imagery




Click on images to see corresponding movies.



Stabilization makes a difference :)
Before and after stabilization :
Row 1: previous frame and sparse flow on it (left) and stabilized frame (right). Row 2: the difference between current frame and (top) previous frame and (bottom) stabilized frame.




Click on images to see corresponding movies.



Stabilized frames, KLT Sparse flow and EM results
Using motion statistics from KLT sparse flow on dense optical flow :
Column 1: stabilized frame (top) and current frame (bottom). Column 2: KLT sparse flow on stabilized frame (top) and motion mask obtained by thresholding the magnitude of the robust flow by KLT motion statistics (bottom). Column 3: horizontal and vertical components of optical flow btw stabilized frame and current frame. Column 4: mixture clusters on sparse flow obtained via EM.




Click on images to see corresponding movies.



Comparison of robust optical flow result with and without pre-stabilization

Comparison of robust optical flow results between frames with no-stabilization and with stabilization:
Column 1 : image sequence and the horizontal and vertical components of robust optical flow results btw successive frames. Column 2 : stabilized frame and the horizontal and vertical components of robust optical flow results after stabilization of previous frame.




Click on images to see corresponding movies.



Experimental results for Sequence-11 : small cars
Results of our approach for a sequence with small cars: Image sequence (top left), magnitude of robust optical flow (bottom left), background ownership weight (top middle), detected vehicles (bottom middle), background mosaic (right).




Click on images to see corresponding movies.



Experimental results for Sequence-13 : zoomed in cars, occlusion and shadows
Results of our approach for a sequence with zoomed in cars, occlusion and shadows: Image sequence (top left), magnitude of robust optical flow (bottom left), background ownership weight (top middle), detected vehicles (bottom middle), background mosaic (right).




Click on images to see corresponding movies.



Experimental results for Sequence-15 : small cars, occlusion and motion parallax
Results of our approach for a sequence with small cars, occlusion and motion parallax: Image sequence (top left), magnitude of robust optical flow (bottom left), background ownership weight (top middle), detected vehicles (bottom middle), background mosaic (right).




Click on images to see corresponding movies.



Experimental results for Sequence-16 : small cars
Results of our approach for a sequence with small cars: Image sequence (top left), magnitude of robust optical flow (bottom left), background ownership weight (top middle), detected vehicles (bottom middle), background mosaic (right).




Click on images to see corresponding movies.



Experimental results for Sequence-20 : Thermal sequence with changing camera gain
Results of our approach for a thermal sequence with changing camera gain: Image sequence (top left), magnitude of robust optical flow (bottom left), background ownership weight (top middle), detected vehicles (bottom middle), background mosaic (right).











[Research]
        [Tracking Vehicles In Airborne Video Imagery]
        [Dense Motion and Appearance Estimation]
        [Implicitization by Matrix Annihilation]
        [Modeling and Measurement Using IPs]
        [Automated Sorting of Remote Controllers ]
        [Shape Assessment by Selective Fixations ]



Related Publications
  • "A Flow-Based Approach to Vehicle Detection and Background Mosaicking in Airborne Video," under review, coming soon. For a technical report, please contact hulyayalcin@gmail.com.

    • Resume | Research | Main Page
    Carnegie Mellon University, Robotics Institute
    5000 Forbes Av., Pittsburgh, PA, 15213
    hulyayalcin@gmail.com