A video tracking algorithm based on multi-template and adaptive feature selection

A feature selection and video tracking technology, applied in computing, image data processing, instruments, etc., can solve problems such as weak single feature expression ability, tracker drift, and redundant features without considering the importance of feature channels. The effect of reducing drift, reducing redundancy, and enhancing robustness

CN109410247AInactive Publication Date: 2019-03-01CHINA UNIV OF PETROLEUM (EAST CHINA)
7 Cites 22 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Publication Date
2019-03-01
Estimated Expiration
Not applicable · inactive patent

Smart Images

  • Figure 1
    Figure 1
  • Figure 2
    Figure 2
  • Figure 3
    Figure 3
Patent Text Reader

Abstract

The invention belongs to the field of computer graphics and image processing, in particular to a video tracking algorithm of multi-template and adaptive feature selection. The invention aims at the single feature problem and the frequent template updating problem of the previous algorithm. A feature fusion method combining convolution feature and manual feature is proposed, and the feature channels are weighted by dimension reduction and entropy value. At the same time, the multi-template is trained to get multi-response map in each frame. After fusion, the final target position and scale arefound, and the template is updated according to the threshold value. The method of the invention has the following beneficial effects: 1. under the condition of ensuring real-time processing, the adaptability of tracking under complex scenes is effectively improved; 2. The accuracy of single target tracking is improved.
Need to check novelty before this filing date? Find Prior Art

Description

technical field

[0001] The invention belongs to the field of computer graphics and image processing, and relates to a video tracking algorithm for multi-template and self-adaptive feature selection. Background technique

[0002] With the advent of the AI ​​era, the new technology of image graphics ushers in a new peak. Single target tracking is an important research direction in the field of machine vision. It is widely used to find the target of interest in the first frame in each frame of a video sequence. In video surveillance, human-computer interaction, medical image processing and other fields. Although some excellent algorithms have emerged in recent years, single-target video tracking is still a challenging task. Various interference factors such as deformation, occlusion, background clutter, and scale changes restrict the improvement of tracking effects.

[0003] Existing algorithms such as CN, KCF, HCF and other algorithms use a single feature to express weakly; S...

Examples

Embodiment Construction

[0039] Below in conjunction with accompanying drawing and specific embodiment the present invention is described in further detail:

[0040] combine figure 1 , a video tracking algorithm with multiple templates and adaptive feature selection, the implementation includes the following steps:

[0041] S1, video sequence preprocessing and neural network pre-training: White balance and histogram equalization are performed on the video sequence. The white balance method is the Gray world method, which aims to reduce the impact of light on the sequence during the tracking process. The training of the neural network uses the imagenet data set to train the VGG-19 network. The convolutional features used in the following steps are the features of the deep convolutional neural network after the convolutional layer. We use the convolutional features after the last convolutional layer. That is conv5-4.

[0042] S2. Initialize the tracking target in the first frame: For the video sequenc...