DETECTING AND TRACKING TRACTOR-TRAILERS USING VIEW-BASED TEMPLATES Masters Thesis Defense by Vinay Gidla Apr 19,2010.

DETECTING AND TRACKING

TRACTOR-TRAILERS USING VIEW-BASED TEMPLATES

Masters Thesis Defense byVinay GidlaApr 19,2010

Introduction• Object tracking: • Sports analysis • Games and gesture recognition • Retail video mining• Automobile driver assistance

• Traffic surveillance• Volume, individual speeds, classification• Lane changes, speed violations,

congestions

Feature-based vehicle tracking

• Beymer et al. 1997 use feature point approach with motion cues to segment vehicles using homography

• Kanhere et al. 2008 use features with 3D estimation using multi-level homographyFeature_based.avi

• Drawback: These approaches track features on the vehicle, not vehicle as a whole

Template-based tracking

• Model the object by 2D template of image intensities

• Compare search image with template image

• Comparison usually by discrete cross-correlation

• Good: Both spatial and appearance informationAble to retrieve shape of the object

• Bad: Encode vehicle appearance from single viewpoint

Do not adapt to changes in appearance of object

Proposal

• Overcome the limitations of a single template by using a template sequence instead of a single template

• The template sequence encapsulates all of the vehicle’s perspective deformations

• As a starting step, aim to detect and track contours of tractor-trailers in multi-lane traffic

Video Sequences

Template creation

Training sequence:

• A portion of traffic video containing a tractor-trailer

• Process the video frames to create a template sequence

Training Sequence

Training frame

Manual contour selection

Template creation

Template sequence

Algorithm Overview

Step 1: Background subtraction

Reference background

Input Video Frame

Background subtracted frame

Step 2: Blob-Template match

Blob-Template match

Plot of Blob-Template match

Step 3: Trace contour

Results based on template-blob correlation

Plot of misalignment

Gradient magnitude match

• Reduce the misalignment by including salient features such as points of high gradient magnitude

• These points are located at identical spatial locations in every tractor-trailer

Training frame

Gradient Magnitude

Template Gradient Magnitude

Template Gradient Sequence

Results based on template-frame gradient

match

Plot of misalignment

Test sequences

Results(Lane 3)

Results(Lane 2)

Level set-based trackingfor automatic template

generation

Conclusion

• The new approach accurately traces the contours of all the tractor-trailers in the traffic video

• Works for multi-lane traffic

• Minor misalignment

Future extensions

• Tracking other classes of vehicles such as passenger cars, buses etc

• Compact template sequence with minimal template redundancy

• Implement matching using level set techniques

Thank You

Questions &

Discussion