Processing Results of Continuous Multi-Obejcts Tracking

using Tensor Voting

 



In this page we show some results obtained with a tensor voting-based tracking. The method performs a perceptual grouping of the detected moving objects.

  • Moving objects are detected using a background-learning techniques
    • Mode-based detection algorithm with background update
    • Shadow extraction algorithm
  • The detected blobs are represented in a graph representation allowing to group moving objects in small clusters in 2D+t
    • Pre-grouping
    • Initial tracking
  • The tensor voting-based technique allows to merge the sub-paths and extract trajectories with consistent motion
  • A sliding-window method with multiple-temporal scales allows to handle incomplet detection, large occlusion and split/merge moving objects

Result 1 ( Two persons crossing : Grouping )

Input Sequence

Detection Result

Tracking Result

 

 

 

Result 2 ( Street : Strong occlusions )

Input Sequence

Detection Result

Tracking Result

 

 

 

Result 3 ( Walking in sweden : Grouping )

Input Sequence

Detection Result

Tracking Result

 

 

 

Result 4 ( One person walking : Complex motion )

Input Sequence

Detection Result

Tracking Result

 

 

 

Result 5 ( Low resolution : False detection )

Input Sequence

Detection Result

Tracking Result

 

 

 

Result 6 ( Two persons crossing & following : Complex motion )

Input Sequence

Detection Result

Tracking Result

 

 

 

Result 7 ( Outdoor sequence : Multiple objects, occlusion, shadows, grouping )

Input Sequence

Detection Result

Tracking Result

 

 

 

Result 8 ( Indoor sequence : Multiple objects, occlusion, shadows, grouping, spliting )

Input Sequence

Detection Result

Tracking Result

 

 

 

Result 9 ( PETS'01 sequence : Multiple objects, grouping, crossing, spliting )

Input Sequence

Detection Result

Tracking Result

 

 

 

Result 10 ( Moving camera : Applying the proposed method after motion compensation )

Detection Result

Tracking Result