CN110782479B

CN110782479B - Visual target tracking method based on Gaussian center alignment

Info

Publication number: CN110782479B
Application number: CN201910947357.3A
Authority: CN
Inventors: 胡钦涛; 毛耀; 周丽君; 周国忠; 何秋农; 周翕; 李志俊; 张超; 乔琦; 聂康
Original assignee: Institute of Optics and Electronics of CAS
Current assignee: Institute of Optics and Electronics of CAS
Priority date: 2019-10-08
Filing date: 2019-10-08
Publication date: 2022-07-19
Anticipated expiration: 2039-10-08
Also published as: CN110782479A

Abstract

The invention provides a visual target tracking method based on Gaussian center alignment, and aims to solve the problem that the target characteristics of the existing target tracking technology cannot be aligned with the modeled Gaussian center. The invention comprises the following steps: step 1, selecting a target initial frame; step 2, extracting target characteristics by using a network; step 3, the target features pass through a center alignment pooling layer; step 4, aligning the result passing through the center alignment pooling layer with the modeled Gaussian center to obtain a minimum loss error; and 5, obtaining a target tracking result. The invention has the beneficial technical effects that: the tracking precision of the Gaussian model can be greatly improved.

Description

Visual target tracking method based on Gaussian center alignment

Technical Field

The invention relates to the technical field of image processing, in particular to a visual target tracking method based on Gaussian center alignment.

Background

The target tracking is widely applied to the fields of automatic driving, aerial photography, video monitoring and the like, most of the traditional image algorithms are characteristic engineering, such as methods of color histograms, HOG characteristics and the like, or methods of particle filtering, Kalman filters and the like, but with the development of the technology, the performance can not meet the current use requirements due to the practical and complex application environment, background similar interference, the change of illumination conditions, shielding and other external factors, target posture change, appearance deformation, scale change, out-of-plane rotation, in-plane rotation, out-of-view, rapid movement, movement blur and the like; with the development of correlation filtering and deep learning, a batch of correlation filtering and deep learning tracking algorithms also appear in succession, from correlation filtering methods proposed by MOSSE algorithm, KCF algorithm, DCF and the like to the deep learning target tracking algorithm such as SiamFC and the like which are developed at a high speed at present, most of the correlation filtering and deep learning tracking algorithms model a target into a gaussian target, however, the target features extracted by the deep learning network or the correlation filter may have the problem of misalignment with the tracked gaussian center (for example, the geometric center point of a human body is near the belly, but the features may be at the head), and meanwhile, it is necessary to develop a tracking algorithm with higher precision and without affecting real-time property.

Disclosure of Invention

The invention aims to provide a visual target tracking method based on Gaussian center alignment. In order to improve the tracking accuracy of the existing tracking algorithm without influencing the real-time performance, the problem that the tracking target feature in the background is not aligned with the Gaussian center of the target to be learned is solved.

In order to solve the technical problems, the invention adopts the following technical scheme:

the visual target tracking method based on Gaussian center alignment comprises the following steps:

step 1, selecting a target initial frame;

step 2, extracting target characteristics by using a network or a related filter;

step 3, passing the target feature through a center alignment pooling layer; the role of the center-aligned pooling layer is to align the network or filter learned features with the modeled gaussian center;

step 4, aligning the result passing through the center alignment pooling layer with the modeled Gaussian center to obtain a minimum loss error; the loss function that is taken is as follows,

wherein, f (x)_j(ii) a w) is the target feature, y_jIs the generated Gaussian model, PRP is the center-aligned pooling layer, γ_jIs the corresponding feature map weight, defined by, for the PRP center pooling layer:

for the target feature

Each element is equal to the sum of its row maximum and column maximum, x_pnIs a row value, x_nqIs the column value.

And 5, obtaining a target tracking result. And obtaining the position of the tracking target according to the obtained response of the target function.

Compared with the prior art, the invention has the following advantages:

1. the invention improves the tracking performance of the original method.

2. The invention is applicable to both the neural network and the related filtering method, and has wide applicability.

3. The invention corrects the Gaussian center by introducing the center alignment pooling layer, has small calculation amount and basically does not change the running speed of the original algorithm in use.

Drawings

FIG. 1 is an overall frame diagram of the present invention;

FIG. 2 is an example of a center-aligned pooling layer of the present invention;

fig. 3 is a comparative example of response gaussian functions with and without the present invention, wherein fig. 3(a) is the gaussian response function of the present invention and fig. 3(b) is the gaussian response function without the present invention.

Detailed Description

The following detailed description of the embodiments of the invention refers to the accompanying drawings. However, the following examples are only for illustrating the present invention in detail and do not limit the scope of the present invention in any way. The programs referred to or relied on in the following embodiments are all conventional programs or simple programs in the art, and those skilled in the art can make routine selection or adaptation according to specific application scenarios.

As shown in fig. 1, the present invention relates to a visual target tracking method based on gaussian center alignment, which comprises the following steps:

step 1, selecting a target initial frame: selecting a target to be tracked, and extracting the characteristics of the target to be tracked;

step 2, extracting target characteristics by using a network or a related filter, such as a ResNet network in the figure 1;

step 3, passing the target feature through a center alignment pooling layer; the role of the center-aligned pooling layer, such as the FIG. 2, is to align the network or filter learned features with the modeled Gaussian center;

step 4, aligning the result passing through the center alignment pooling layer with the modeled Gaussian center to obtain a minimum loss error; the loss function to be taken is as follows,

wherein, f (x)_j(ii) a w) is the target feature, y_jIs the generated Gaussian model, the last Gaussian in FIG. 1, that is, the Gaussian in FIG. 3(b), PRP is the center-aligned pooling layer, γ_jIs the corresponding feature map weight, defined by, for the PRP center pooling layer:

for the target feature

Fig. 3 is a comparison of gaussian plots generated using and without the present invention, where fig. 3(a) is a gaussian response plot of the present invention, which can be found to be smoother than fig. 3(b), and to have only one gaussian response.

In order to verify the effectiveness of the invention, ATOM and Simmask algorithms are selected as comparative examples, a standard VOT2018 data set is used for comparing results, and the precision A and robustness R of the tracking method and the expected average coverage EAO are compared.

Table one: comparison of results on data sets for examples and comparative examples

Method	A	R	EAO
				ATOM	0.590	0.204	0.401
Siammask	0.609	0.276	0.380
				The invention	0.612	0.169	0.434

As can be seen from the table I, compared with the prior advanced ATOM and Simmask, the method of the invention improves the precision A (the higher the better), the robustness R (the lower the better) and the expected average coverage EAO (the higher the better).

Claims

1. A visual target tracking method based on Gaussian center alignment is characterized by comprising the following steps:

step 1, selecting a target initial frame;

step 2, extracting target characteristics by using a network;

step 3, the target features pass through a center alignment pooling layer;

step 4, aligning the result passing through the center alignment pooling layer with the modeled Gaussian center to obtain a minimum loss error;

step 5, obtaining a target tracking result;

the step 3 of aligning the center of the pooling layer is to re-correct the target feature from the eccentricity of the object to the center of the object;

the loss error function adopted in step 4 is:

wherein, f (x)_j(ii) a w) is the target feature, y_jIs a generated Gaussian model, PRP is a center pairQuasi-pooling layer, γ_jIs the corresponding feature map weight;

for PRP central pooling layer, by the following definition:

for the target feature