Online learning of robust object detectors during unstable tracking

ArticleThis section describes the TMD-tracing model detection tracking system.

This article mainly introduces online learning.AlgorithmOr a system. I personally think This is the prototype of the Zdenek PN-learning algorithm.

This online learning algorithm consists of two parts.

1 growing events select a positive sample and add it to the training set.

2 pruning events trim negative samples to make the model more lubang.

In fact, the so-called model in this article is just a series of Target Detection windows. (Similar to the eigenface model ).

The author assumes that:

Let's assume there is a set of L *, L * indicates all coordinate regions where the target may exist, and L * is the correct target in the middle (in a rational condition)

Our detection model is assumed to be <, lt is just a hypothetical model, and we hope that lt can be infinitely close to L *.

Then growing events is constantly trying to add the target certificate sample to lt.

And pruning events is to pick out the negative samples in lt without a segment.

The final goal is to make lT = L *.

This is the TMD Training Principle. So how to add and trim positive and negative samples?

Growing events follows three principles to add positive samples:

1 ABS (absolute distance from first patch) is actually to make the possible sample and the first in LT, that is, the initial target feature comparison distance.

2 diff (difference between consecutive patches) This is possible to target a de-compare distance from a feature of the latest Xt-1.

3 Loop: a loop process adds the target of a positive sample to Lt, which greatly differs from lt, to improve the robustness of the model.

Pruning events indicates trimming the negative samples in LT, that is, the previous judgment is for positive samples, and with the increasing robustness of LT, the current judgment is for negative samples. At this time, we will remove these negative samples.

In fact, during implementation, the author defines the proportion of a negative sample, as long as the proportion is exceeded, we all take negative samples and remove them.

Other implementation details:

Tracking: optical flow is used. Select a feature point for tracking.

Online Model: A method similar to eigenface.

Features: 2bit biary patterns, a feature similar to the local code and Haar.

References:

Http://info.ee.surrey.ac.uk/Personal/Z.Kalal/tld.html

Original article address:

Http://info.ee.surrey.ac.uk/Personal/Z.Kalal/Publications/2009_olcv.pdf