On-line fusion of trackers for single-object tracking

On-line fusion of trackers for single-object tracking

Pattern Recognition, 2018 - Elsevier

2019-08-18 22:31:01

Paper: https://www.sciencedirect.com/science/article/pii/S0031320317303783

1. Background and Motivation:

As we all known, regular single object trackers are easily influenced by chanllenging factors and NO single tracker can handle all these factors well. And different trackers may works well under different scenario, therefore, how to fuse existing trackers to achieve robust tracking is a worthy studying research topic, right? The authors classified existing multi-tracker fusion based algorithms into two main categoreis:

1). passive fusion: only combine trackers outputs with no interaction between the trackers.

2). active fusion: integrate data provided by each tracker with the objective of correcting their inner model when necessary.

In addition, the authors also classified existing multi-tracker fusion techniques into the following three kinds:

The authors state that the active fusion leads in general to better performance, but necessitates a control over tracker components and update mechanisms. This paper inroduce a complementarity measure between trackers based on individual drift measures to predict the fusion performance of the combined trackers in order to select it.

2. Offline tracker evaluation.

The first thing before tracking fusion is to evaluate the tracking performance of each tracker, then, we can design novel strategy to fuse them. The authors propose two kinds of evaulate methods, i.e. the gobal evaluation and local evaluation method:

2.1 Global evaluation.

In this section, the authors only simply give an introduction about evaluation metric of VOT challenging competition, i.e. the accuracy and robustness.

2.2 Local evaluation.

In addition to the global evaluation, the authors also introduced a fine-grained local evaluation method, named "incompleteness".

Incmpleteness is used to define the inability of the trackers to compensate collectively for drifting, and is computed as the number of times when all trackers are simultaneously drifting at the same time (所有跟踪算法同时失效的次数). Formally, the incompleteness I of a set of M trackers on a database of N frames as:

where the $d_t^i$ is the variable used to indicate the tracker $T_i$ is drifting or not.

3. Online tracker failure prediction.

The authors attempt to predict tracking failures from a set of M parallel trackers T = [T1, T2, ... , TM], either individually or collectively. They use three ways to estimate the tracking failure.

3.1 Behavioral Indicators (BI) 

They consider three kinds of information from used trackers, i.e. the confidence score, the score map and specific indicators.

confidence score: this is a popular used criterion to measure the tracker is drift or not. Because they assume the score will be high, when the tracker works well, but rather low when failure.

score map: the tracker usually predict their bounding box based on this response map.

specific indicator: designed for more complicated trackers.

3.2 Box Filtering (BF) 

When the current estimated location of the target from tracker is very far from the previous estimated location output by fusion.

3.3 Box Consensus (BC) 

The principle of this criterion is they think: only few trackers in a given collections are likely to drift. They think the outlier is the failed tracker.

4. Proposed Fusion Method 

如上图所示，作者将整个跟踪过程分为四个阶段：同时进行多跟踪器的跟踪，跟踪器选择，跟踪器融合，跟踪器的校正。

4.1 Tracker parallel running:

就是同时跑多个跟踪算法；

4.2 Tracker selection by on-line failure prediction:

从上述跟踪算法的结果中，进行 failure 的预测，然后选择那些高置信度的结果。

4.3 Fusion bounding box computation

在拿到所要融合的 Bbox 之后，作者用如下两种方法进行融合：

1）平均处理：即，将多个 BBox 的坐标进行平均，融合为一个结果。

2）Center of gravity (Gray)：加权 k个 box 。

6.4 Tracker correlation:

作者提出了三种方法来校正跟踪模型：

==