NEWS CENTER
Position :
Homepage
/
/
/
Briefly describe the visual tracking algorithm

Briefly describe the visual tracking algorithm

  • Categories:Company news
  • Author:
  • Origin:
  • Time of issue:2019-07-19
  • Views:0

(Summary description) In recent years, visual tracking, as a basic problem in computer vision, has been widely studied. Tracking algorithms are also used in a wide range of applications, such as driverless cars and our smart luggage.

In recent years, visual tracking, as a basic problem in computer vision, has been widely studied. Tracking algorithms are also used in a wide range of applications, such as driverless cars and our smart luggage.

Briefly describe the visual tracking algorithm

(Summary description) In recent years, visual tracking, as a basic problem in computer vision, has been widely studied. Tracking algorithms are also used in a wide range of applications, such as driverless cars and our smart luggage.

In recent years, visual tracking, as a basic problem in computer vision, has been widely studied. Tracking algorithms are also used in a wide range of applications, such as driverless cars and our smart luggage.

  • Categories:Company news
  • Author:
  • Origin:
  • Time of issue:2019-07-19
  • Views:0
Information

【 Abstract 】 : In recent years, visual tracking, as a basic problem in computer vision, has been widely studied. Tracking algorithms are also used in a wide range of applications, such as driverless cars and our smart luggage.

In recent years, visual tracking, as a basic problem in computer vision, has been widely studied. Tracking algorithms are also used in a wide range of applications, such as driverless cars and our smart luggage.

The main function of the visual tracking algorithm is that in the video sequence, given a rectangular box of a target, the algorithm continuously locates the target position in the subsequent video sequence.

How to effectively use machine learning methods to solve the tracking problem and how to transform the problem is a very core problem. In general, tracking algorithms are divided into generative tracking algorithm and discriminant tracking algorithm. I'll look at some of the finer distinctions below.

Classifier:

Obviously, for the tracking problem, we can construct a classifier of foreground (target) -background classification. By sampling a large number of positive samples around the target and a large number of negative samples around the far target, a training set of foreground - background samples can be formed. On this basis, we can use SVM, tree based discriminant model for classification.

The disadvantages of this approach are also very obvious. In order to obtain a classifier with strong discriminability, we need to extract a large number of positive and negative samples. And for these samples, it will take more time to extract features. In addition, discriminator learning is also very slow. This makes the real-time application of tracking algorithm restricted.

Correlation filtering (linear SCORE regression)

For a long time, visual tracking algorithms are based on the above particle filter framework. Until 2010, Bolme introduced correlation filtering into visual tracking. This algorithm USES the cyclic shift sampling method to sample the target image intensively. At the same time, the algorithm transforms the time domain problem into the frequency domain by means of Fourier transform, and obtains more than 600 FPS(Frame per second).

On this basis, nuclearization and multi-channel expansion are carried out to obtain the familiar KCF algorithm. From 2013 to 2017, it can be said that this algorithm flourished, and basically dominated the target tracking field (more than half of the target tracking algorithms are based on correlation filtering every year).

In addition to the above development background, we also need to point out that, compared with the previous classifier algorithm, this algorithm only USES the cyclic shift sampling method in the way of sampling, and converts the 0-1 problem of classification into a gaussian response.

Position return

In the above context, while we focus on tracking the development of the field, we also see the development of other fields simultaneously. Since the birth of AlexNet in 2012 NIPS conference, deep learning has announced a full recovery. Up to now, deep learning is almost an unavoidable tool when we discuss approaches in various fields.

In the same way, deep learning is gradually sweeping the field of target tracking. A large number of algorithms have significantly improved their performance by directly replacing traditional manual features with deep learning features. In this section, we mainly explain GOTURN, a deep learning algorithm based on location regression. This is also the first real-time end-to-end deep learning tracking algorithm. This algorithm conducts a lot of training and learning in the offline process. In the online process, the search image and search area are given directly, and the target's position in the search image is directly returned through the neural network.

The breakthrough with this algorithm is that you don't need anything in the online process, right? Netuning and, unlike most methods based on classification and correlation filtering, direct location regression. This also allows the algorithm to achieve a high efficiency of 100FPS on GTX1060 without the need for a large number of samples.

Decisions (RL)

 睿牛机器人

In addition to the above method of position regression, we can also transform the tracking problem into a decision problem. When given a target image and a search area, we can choose whether to adjust the image up, down, left, right, or zoom in and out (each adjustment is considered a decision).

When the problem is transformed into a decision problem, we can use supervised learning and reinforcement learning (RL) for training. At the same time, compared with the traditional sampling-based method, this algorithm only needs several iterations to correct and find the target, instead of sampling a large number of samples, which is also more efficient than the traditional algorithm.

By transforming the tracking algorithm in several ways, we can see that a simple tracking problem, with different problem definitions, can produce so many changes. This also proves that there are a thousand Hamlets in a thousand readers' minds. It's a way of creating by looking at things differently.

Scan the QR code to read on your phone

 RENUE

Weld tracking is no worry,     Go to renue when in trouble!

ADD:Haosheng Industrial Park, Kunshan, Suzhou
TEL:
0512-80171802  /  0512-80171805
PHONE:MR. 18626329319
E-MAIL:zhangxiangfeng@renue.com.cn
FAX:0512-80171801
WEB:
www.renue.com.cn

Copyright ©  2020   Suzhou Renue Robot Technology Co.,Ltd.  All rights reserved     苏ICP备25222523号  

Search