Online Visual Object Tracking Using Incremental Discriminative Color Learning

Alireza Asvadi, Hami Mahdavinataj, Mohammadreza Karami, and Yasser Baleghi
Babol University of Technology


Object model is one of the essential components for visual tracking. Generally, tracking algorithms use generative or discriminative methods to model the object appearance. Generative methods build a reference model of an object and use this model to search for the object in the next frames. On the other hand, discriminative methods distinguish an object from surrounding background in every frame without using a definitive model of the object. These methods are based on training a classifier to distinguish the object from surrounding background. However, discriminative methods do not explicitly model object. Therefore, noisy samples are likely to interfere and cause visual drift.
This paper presents a method for tracking an object in a sequence of images given its location in the first frame. In this paper, a combination of generative and discriminative methods is used to model the object appearance. It is used to provide a right balance between being adaptive and ability to re-track the object after losing the object due to occlusion. The quantized 3D joint RGB histograms of the region within the inner rectangle and the region between the inner and outer rectangles are used as a discriminative component. The generative component is composed of the positive part of the log-likelihood ratio of the computed 3D joint RGB histograms along with an incremental color learning scheme with a forgetting factor. In every frame, after localization, discriminative component differentiates the object from its surrounding background and adds the detected object colors to the generative model. The generative model is evolved during tracking and is used to detect the object in the next frame.
It is shown the proposed method can handle visual drift effectively. Evaluated against five state of the art methods, experiments demonstrate superior results of the tracking algorithm. Implemented in MATLAB, the algorithm runs at 17.2 frames per second, including image input/output time.


Click [Results] to download all video results in a zip format. [Youtube link]

Effect of the model update procedure on the object model and corresponding confident map


The tracker has been tested on variety of challenging sequences, and twelve of the most representative sequences are reported in this paper. The sequences are Basketball , Crossing , David outdoor , Girl , Face occlusion, Tiger and Trellis available at Tracker Benchmark v1.0, Sunshade and Torus available at VOT2013 dataset, Caviar1 and Caviar2 from Caviar dataset and Board from Prost Dataset.

Evaluated Methods:

The proposed method (Incremental Discriminative Color Tracking called IDCT) is evaluated against five methods which their codes are publicly available, Mean Shift tracker (MS), mean-shift tracker with Corrected Background Weighted Histogram (CBWH), Variance Ratio tracker (VR), Fragment based Tracker (Frag) and Locality Sensitive Histogram Tracker (LSHT).


2.  A. Asvadi, H. Mahdavinataj, M. Karami and Y. Baleghi, “Online Visual Object Tracking Using Incremental Discriminative Color Learning” CSI Journal on Computer Science and Engineering (JCSE), vol. 12, no. 2 & 4 (b), pp. 16-28, 2014.
[link to PDF]

1. A. Asvadi, H. Mahdavinataj, M. Karami, and Y. Baleghi, “Incremental Discriminative Color Object Tracking” Artificial Intelligence and Signal Processing, Communications in Computer and Information Science, ISBN 978-3-319-10849-0, Volume 427, Springer-Verlag 2014, pp 71-81. DOI:10.1007/978-3-319-10849-0_8
[link to PDF] [presentation] [code] [result]


The code runs on Windows XP with MATLAB R2011a.