Appearance management and cue fusion for 3D model-based tracking

Publisher Website
Google Scholar
Abstract
This paper presents a systematic approach to acquir- ing model appearance information online for monocular model-based tracking. The acquired information is used to drive a set of complementary imaging cues to obtain a highly discriminatory observation model. Appearance is modeled as a Markov random field of color distributions over the model surface. The online acquisition process esti- mates appearance based on uncertain image measurements and is designed to greatly reduce the chance of mapping non-object image data onto the model. Confidences about the different appearance driven imaging cues are estimated in order adaptively balance the contributions of the differ- ent cues which allows to maintain performance in the pres- ence of degradation in imaging conditions. The discrim- inatory power of the resulting model is good enough to allow long-duration single-hypothesis model-based track- ing under flexible imaging conditions with no prior appear- ance information. The performance of the resulting model- based tracker is evaluated carefully based on real and semi- synthetic video sequences, showing that the presented algo- rithm is able to robustly track a vide variety of targets under challenging conditions. the true model parameters can be conducted with one or at least very few active hypotheses. The key to powerful observation models is object specific appearance, which al- lows direct model to image matching and precise discrimi- nation between foreground and background image regions. We follow the approach to continuously acquire appearance information from image data directly during tracking. This process has to be done carefully because mapping back- ground data to the model can quickly destabilize the tracker. Building on previous work (7), we present in this paper a tracking framework based on robust and reliable appear- ance model acquisition and maintanance. The discrimina- tory power of the resulting model is good enough to allow long-duration single-hypothesis model-based tracking with no prior appearance information. In addition, by performing confidence guided cue-fusion, the tracker is able to adapt to changing environmental conditions. When some of the vi- sual cues become degraded, confidence measures rebalance the contribution of other cues allowing stable and robust tracking of general targets under a wide range of challeng- ing operating conditions.

This publication has 9 references indexed in Scilit: