Kinematic object track stitcher for post tracking fragmentation detection and correction

Date of Award


Degree Name

M.S. in Electrical Engineering


Department of Electrical and Computer Engineering


Advisor: Eric John Balster


A common goal of tracking objects over multiple image frames is to ensure that the object tracks are as accurate and error free as possible. One of the biggest impediments to accurate object tracking is an event called track fragmentation. Track fragmentation occurs when an object's trajectory or track is reported by the tracker as two or more separate track instances. Fragmentation occurs because of features present in the imagery and the method at which the imagery is processed. Some causes of track fragmentation include: occlusions, object trajectory or velocity changes, proximity of objects to one another, parallax, image stitching, lighting changes, and image artifacts. Track fragmentation is tackled in multiple ways. The GATER tracker uses projected points to make predictions where tracks will be in future frames to compensate for events that cause fragmentation. Another solution is to use a process known as track stitching. A post-tracking kinematic track stitcher is proposed that is able to detect and correct track fragmentation, improving fragmentation from GATER's projected points. The proposed track stitcher requires only a track's centroid and frame, information any tracker can provide. The track stitcher will improve the accuracy of the tracks by reducing track fragmentation. The track stitcher consists of a series of filters, including a temporal filter and multiple kinematic filters, designed to determine the likelihood and feasibility of a track being fragmented. The filters are designed to be easily enabled, disabled, and integrated depending on the needs of a given data set. These filters include sets of parameters that can be manually altered by the user for different tracking scenarios. A selection algorithm, in this case Munkres, is used to evaluate the track scores and make stitches accordingly. The track stitcher is run on multiple track sets acquired from a Wide Area Motion Imagery image set. For the first experiment, one of the image sets is selected for tuning, and then the track stitcher is run with varying parameters for the kinematic filter and scoring filter. The best performing parameters for the first image set are then used to test the track stitcher on the remaining image sets in order to prove that the same parameters can be universally applied to similar imagery. In every test case the track stitcher reduces the amount of track fragmentation and error (spuriousness) present in the tracks. In a majority of the tests, the purity of the tracks also increases. When compared to the point projected tracks GATER produces, the proposed track stitcher improves spuriousness by 26.67% and fragmentation by 17.67% on average.


Image registration, Figure-ground perception, Automatic tracking, Pattern recognition systems, Electrical Engineering, Track stitching, GATER, Track fragmentation, Object tracking, Projected points

Rights Statement

Copyright 2015, author