Parameter estimation and contextual adaptation for a multi-object tracking CRF model

We present a detection-based approach to multi-object tracking formulated as a statistical labeling task and solved using a Conditional Random Field (CRF) model. The CRF model relies on factors involving detection pairs and their corresponding hidden labels. These factors model pairwise position or color similarities as well as dissimilarities, and one critical issue is to be able to learn their parameters in an accurate and unsupervised way. We argue in this paper that tracklets and local context can help to obtain relevant parameters. In this context, the contributions are as follows: i) a factor term global parameter estimation based on intermediate tracking results; ii) a detection dependent parameter adaptation scheme that allows to take into account the local detection contextual information during online tracking. Experiments on PETS 2009 and CAVIAR datasets show the validity of our approach, and similar or better performance than recent state-of-the-art algorithms.