CN110135365B - Robust object tracking method based on hallucination adversarial network - Google Patents

Abstract

Robust target tracking method based on hallucination adversarial network, involving computer vision technology. First, a new hallucination adversarial network is proposed, which aims to learn the nonlinear deformation between sample pairs, and apply the learned deformation to the new target to generate new target deformation samples. In order to effectively train the proposed hallucination adversarial network, a deformation reconstruction loss is proposed. Based on the hallucination adversarial network trained offline, a target tracking method based on the hallucination adversarial network is proposed, which can effectively alleviate the overfitting problem of the deep neural network due to online update in the target tracking process. In addition, in order to further improve the quality of deformation transfer, a selective deformation transfer method is proposed, which further improves the tracking accuracy. The proposed object tracking method achieves competitive results on current mainstream object tracking datasets.

Description

Robust object tracking method based on hallucination adversarial network

technical field

The invention relates to computer vision technology, in particular to a robust target tracking method based on hallucination confrontation network.

Background technique

In recent years, the application of deep neural networks in the field of computer vision has achieved great success. As one of the basic problems in the field of computer vision, object tracking plays a very important role in many current computer vision tasks, such as unmanned driving, augmented reality, robotics and other fields. Recently, the research of target tracking algorithm based on deep neural network has received extensive attention from researchers at home and abroad. However, unlike other computer vision tasks (such as object detection and semantic segmentation), the application of deep neural networks in object tracking tasks is still very effective. Therefore, the generalization of deep neural networks is greatly limited, which in turn affects the tracking results. At the same time, the target tracking task is designed to track any target, and it does not give any prior knowledge of the target to be tracked in advance, which also brings great challenges to the selection of offline training datasets for deep neural networks. Therefore, it is of great practical significance to propose a target tracking algorithm based on deep neural network with strong generalization.

In order to alleviate the above problems, researchers at home and abroad have mainly proposed two types of solutions. The first method regards the target tracking task as a template matching problem, and its specific implementation often uses a deep siamese network, taking the target template and the search area as the input of the deep siamese network at the same time, and finally obtains the search area that is most similar to the target template. sub-region location. The deep Siamese network based on similarity calculation can be trained completely offline by using a large number of labeled target tracking data sets, so it can avoid the overfitting problem caused by too few online training samples. In the target tracking algorithm based on deep Siamese network, its pioneering algorithm is SiamFC. Based on SiamFC, researchers have proposed many improved algorithms, including SiamRPN using region proposal window generation network, MemSiamFC using dynamic memory network, SiamRPN++ using deeper skeleton network, etc. Since SiamFC-type tracking algorithms avoid time-consuming online training steps, they often achieve tracking speeds far beyond real-time. However, since such algorithms lack an online learning process for target appearance changes, their accuracy is still limited (such as the accuracy results on the OTB dataset). Another class of methods proposed by researchers aims to learn robust neural network classifiers with limited online samples. The general idea of such methods is to use methods in the field of transfer learning to alleviate the overfitting problem, and a more representative method is MDNet proposed by H. Nam et al. in 2016. MDNet first uses multi-domain offline learning to learn better initial model parameters of the classifier, and then further trains the classifier by collecting positive and negative samples of the target during the tracking process. Recently, based on MDNet, researchers have proposed VITAL using adversarial learning, BranchOut learning target representations at different levels, SANet using RNN, etc. Compared with the previous method, this method can achieve higher tracking accuracy than the previous method. However, due to the extremely limited online samples (especially target samples), the online learning of such methods is very limited, and it is still prone to overfitting, which in turn affects the tracking performance. Therefore, it is of great significance to design a simple and effective method to alleviate the overfitting problem of deep target tracking algorithms during the tracking process.

Compared with current object tracking algorithms, humans can track moving objects with ease. Although the mechanism of the human brain has not been fully explored so far, we can be sure that through the previous learning experience of human beings, the human brain has derived an unparalleled imagination mechanism. Humans can learn similar actions or transformations from various things they see at ordinary times, so as to apply such similar transformations to different targets, so as to imagine the appearance of new targets in different poses or actions. Such an imaginary mechanism is very similar to the data augmentation method in machine learning. The human brain can be compared to a visual classifier, and then the imaginary mechanism is used to obtain target samples in different states, thereby training a robust visual classifier.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a robust target tracking method based on hallucination confrontation network.

The present invention includes the following steps:

1) Collect a large number of deformation sample pairs in the labeled target tracking data set as a training sample set;

然后在后20帧内随机选取一帧中的目标样本作为

用于构成一组形变样本对

选取大量的形变样本对构成训练样本集合；所述数据集为Fei-Fei Li等人在2015年提出的ILSVRC-2015视频目标检测数据集。In step 1), the specific process of collecting a large number of deformation sample pairs in the marked target tracking data set as a training sample set may be: marking a video sequence to collect a large number of target sample pairs, and a pair of samples contains the same target; In a, first select the target sample in the t-th frame

Then randomly select the target sample in one frame in the next 20 frames as

used to form a set of deformed sample pairs

A large number of deformed sample pairs are selected to form a training sample set; the data set is the ILSVRC-2015 video target detection data set proposed by Fei-Fei Li et al. in 2015.

2) Feature extraction is performed on all samples in the training sample set obtained in step 1) to obtain a training sample feature set;

In step 2), the feature extraction step may be as follows: first, use the bilinear interpolation method to change the size of the target sample to 107×107×3, and then use the neural network feature extractor Feature extraction is performed on the target sample; the structure of the feature extractor φ(·) can be the first three convolutional layers of the VGG-M model pre-trained on the Imagenet dataset.

3) using the training sample feature set, adversarial loss and the proposed deformation reconstruction loss obtained in step 2) to train the proposed hallucination adversarial network offline;

和

使用幻觉对抗网络学习

和

间的形变，并将此形变施加到

用以生成关于目标b新的形变样本，使用对抗损失保证生成的样本分布与目标b分布相近：In step 3), the training process can be as follows: first, two sets of training sample feature pairs are selected from the training sample feature set, which are expressed as

and

Learning with Illusion Adversarial Networks

and

deformation between and apply this deformation to

It is used to generate new deformed samples about the target b, and the adversarial loss is used to ensure that the generated sample distribution is similar to the distribution of the target b:

E_n和D_e分表表示所提出的对抗幻想器中的编码器和解码器部分；为了使得生成样本有效编码形变z^a，提出形变重构损失对生成样本进行约束：in,

_En and D _e sub-tables represent the encoder and decoder parts in the proposed adversarial fantasy device; in order to make the generated samples effectively encode the deformation za , ^a deformation reconstruction loss is proposed to constrain the generated samples:

最终，用于离线训练所提出的幻觉对抗网络的总损失函数为：in,

Finally, the total loss function for offline training of the proposed hallucination adversarial network is:

l _all =l _adv +λl _def , (Formula 3)

where λ is the hyperparameter used to balance the two losses;

The offline training of the hallucination confrontation network may include the following sub-steps:

3.1 The parameter λ in formula (3) is set to 0.5;

3.2 In training, the optimizer used is Adam (DPKingma, and JLBa, “Adam: A method for stochastic optimization,” in Proceedings of the International Conferenceon Learning Representations, 2014), the number of iterations is 5×10 ⁵ , and the learning rate is 2 × ^10-4 ;

3.3 The encoder and decoder structures of the proposed hallucination adversarial network are both three-layer perceptrons with 2048 hidden layer nodes, 9216 encoder input layer nodes, 64 encoder output layer nodes; decoder input layer nodes are 4672; the discriminant network is also a three-layer perceptron with 2048 hidden layer nodes, the number of input nodes is 9216, and the number of output nodes is 1.

4) Given the first frame annotated image in the test video, collect target samples, and use Gaussian and random sampling methods to sample positive and negative samples around the target samples;

In step 4), the details of the sampling can be: in each iterative training, the ratio of positive and negative samples is sampled at a ratio of 1:3, that is, 32 positive samples and 96 negative samples, and the positive sample determination standard is The regional overlap ratio of the sampled sample and the target sample is greater than 0.7, and the criterion for negative samples is that the regional overlap ratio of the sampled sample and the target sample is lower than 0.5.

5) Use the proposed selective deformation migration method to select the pair of samples to be migrated for the tracking target;

表示视频片断s_i中对应样本的个数；对于视频片断s_i的特征表达ψ(s_i)，可以通过如下方式计算得到：In step 5), the process of selecting the pair of samples to be migrated may be: defining N _s to represent the number of video clips in the data set used to collect deformation sample pairs, and s _i to be the identification of the video clips, wherein,

Represents the number of corresponding samples in the video segment s _i ; for the feature expression ψ(s _i ) of the video segment s _i , it can be calculated as follows:

为深度特征提取器，对于目标特征

计算其余每个视频片断表征ψ(s_i)间的欧式距离，选取距离最近的T个视频片断；在选择的T个视频片断中，采用与步骤1)中相同的方式收集大量的形变样本对，构成集合D_S，用于后续目标形变迁移；in,

is the deep feature extractor, for the target feature

Calculate the Euclidean distance between the remaining video clips representing ψ(s _i ), and select the T video clips with the nearest distance; in the selected T video clips, collect a large number of deformation sample pairs in the same way as in step 1). , forming a set D _S for subsequent target deformation migration;

The selective deformation transfer method may include the following sub-steps:

为去掉全连接层的ResNet34模型；5.1 The deep feature extractor used when computing the feature representation of video clips

To remove the ResNet34 model of the fully connected layer;

5.2 When selecting similar video clips, the parameter T is set to 2×10 ³ .

6) Based on the selected sample pairs to be migrated, use the offline trained hallucination confrontation network to generate deformed positive samples;

In step 6), based on the selected sample pairs to be migrated, the specific steps of using the offline trained hallucination confrontation network to generate deformed positive samples may be: in each iteration training, randomly select 64 samples from the set D _S For sample pairs, each pair of samples is input against the target sample to generate corresponding deformation samples, and finally, for each iteration, a total of 64 positive samples are generated.

7) Use the positive and negative samples of spatial sampling and the generated positive samples to jointly train the classifier, and the resulting classification error loss is used to update the classifier and the hallucination confrontation network at the same time;

In step 7), the classifier is jointly trained using the positive and negative samples of spatial sampling and the generated positive samples, and the classification error loss generated by the classifier is used to update the classifier and the hallucination confrontation network at the same time. The specific method can be: The generated 64 positive samples, 32 spatially sampled positive samples, and 96 spatially sampled negative samples are input to the classifier, and the cross-entropy loss of the binary classification is calculated, and then the Adam optimizer is used to update the classifier at the same time through the back-propagation algorithm. And the hallucinations adversarial network.

8) Given a new test frame, use the region with the highest confidence of the trained classifier as the target position to complete the tracking of the current frame;

In step 8), the given new test frame, using the region with the highest confidence of the trained classifier as the target position, the specific process of completing the tracking of the current frame can be: in the current test frame, use random sampling at the same time With Gaussian sampling, sample sampling is performed at the target position estimated in the previous frame; the sampled sample is input to the classifier to obtain its corresponding target confidence.

The invention aims to apply the imagination mechanism of the human brain to the current target tracking algorithm based on deep learning, and proposes a new robust target tracking method based on the hallucination confrontation network. The present invention first proposes a new hallucination confrontation network, which aims to learn the nonlinear deformation between pairs of samples, and apply the learned deformation to a new target to generate a new target deformation sample. In order to effectively train the proposed hallucination adversarial network, a deformation reconstruction loss is proposed. Based on the hallucination adversarial network trained offline, a target tracking method based on the hallucination adversarial network is proposed, which can effectively alleviate the overfitting problem of the deep neural network due to online update in the target tracking process. In addition, in order to further improve the quality of deformation transfer, a selective deformation transfer method is proposed, which further improves the tracking accuracy. The target tracking method proposed by the present invention has achieved competitive results on the current mainstream target tracking data sets.

Description of drawings

FIG. 1 is a schematic flowchart of an embodiment of the present invention.

Detailed ways

The method of the present invention will be described in detail below in conjunction with the accompanying drawings and examples. The present example is implemented on the premise of the technical solution of the present invention, and the implementation manner and specific operation process are given, but the protection scope of the present invention is not limited to the following example.

Referring to Fig. 1, the embodiment of the present invention includes the following steps:

然后在后20帧内随机选取一帧中的目标样本作为

以此来构成一组形变样本对

按照上述步骤，选取大量的形变样本对来构成训练样本集合。A. Collect a large number of deformation sample pairs in an annotated target tracking dataset as a training sample set. The specific process is as follows: label the video sequence to collect a large number of target sample pairs (a pair of samples contains the same target). As in the video sequence a, first select the target sample in the t-th frame

Then randomly select the target sample in one frame in the next 20 frames as

This constitutes a set of deformation sample pairs

According to the above steps, a large number of deformation sample pairs are selected to form a training sample set.

B. Perform feature extraction on all samples in the training sample set obtained in step A to obtain a training sample feature set. The feature extraction steps are as follows: first, the target sample is changed to 107×107×3 using the bilinear interpolation method, and then the feature extraction is performed on all the interpolated target samples by using the neural network feature extractor φ(·).

和

使用幻觉对抗网络学习

和

间的形变，并将此形变施加到

用以生成关于目标b新的形变样本。使用对抗损失保证生成的样本分布与目标b分布相近：C. Use the training sample feature set obtained in step B, the adversarial loss, and the proposed deformation reconstruction loss to train the proposed hallucination adversarial network offline. The specific training process is described as follows: First, two sets of training sample feature pairs are selected from the training sample feature set, expressed as

and

Learning with Illusion Adversarial Networks

and

deformation between and apply this deformation to

Used to generate new deformation samples about target b. Using an adversarial loss ensures that the generated sample distribution is close to the target b distribution:

E_n和D_e分表表示所提出的幻觉对抗网络中的编码器和解码器部分。为了使得生成样本有效编码形变z^a，提出形变重构损失对生成样本进行约束：in,

The _En and _De sub-tables represent the encoder and decoder parts in the proposed hallucination adversarial network. In order to make the generated samples effectively encode the deformation za, ^a deformation reconstruction loss is proposed to constrain the generated samples:

最终，用于离线训练所提出的幻觉对抗网络的总损失函数为：in,

Finally, the total loss function for offline training of the proposed hallucination adversarial network is:

l _all =l _adv +λl _def , (Formula 3)

where λ is the hyperparameter used to balance the two losses.

D. Given the first frame annotated image in the test video, collect target samples, and use Gaussian and random sampling methods to sample positive and negative samples around the target samples. The sampling details are described as follows: In each iteration of training, the ratio of positive and negative samples is sampled in a ratio of 1:3, that is, 32 positive samples and 96 negative samples. The criterion for positive samples is that the regional overlap ratio between the sampled sample and the target sample is greater than 0.7, and the criterion for negative samples is that the regional overlap ratio between the sampled sample and the target sample is less than 0.5.

表示视频片断s_i中对应样本的个数。对于视频片断s_i的特征表达ψ(s_i)，可以通过如下方式计算得到：E. Use the proposed selective deformation transfer method to select pairs of samples to be migrated for the tracking target. The specific selection process is described as follows: define N _s to represent the number of video clips in the data set used to collect deformation sample pairs, s _i to be the identity of the video clips, where,

Indicates the number of corresponding samples in the video segment _si . For the feature expression ψ(s _i ) of the video segment _si , it can be calculated as follows:

为深度特征提取器。对于目标特征

计算其余每个视频片断表征ψ(s_i)间的欧式距离，选取距离最近的T个视频片断。在选择的T个视频片断中，采用与步骤A中相同的方式收集大量的形变样本对，构成集合D_S，用于后续目标形变迁移。in,

is a deep feature extractor. For target features

Calculate the Euclidean distance between the representation ψ(s _i ) of each of the remaining video clips, and select the T video clips with the closest distance. In the selected T video clips, a large number of deformation sample pairs are collected in the same manner as in step A to form a set D _S for subsequent target deformation migration.

F. Based on the selected sample pairs to be migrated, use the offline trained hallucination confrontation network to generate deformed positive samples. The specific generation steps are as follows: in each iterative training, 64 pairs of samples are randomly selected from the set D _S , and each pair of samples is input against the target sample to generate corresponding deformation samples. Finally, for each iteration, a total of 64 positive samples are generated.

G. The classifier is jointly trained using the spatially sampled positive and negative samples and the generated positive samples, and the resulting classification error loss is used to simultaneously update the classifier and the hallucination adversarial network. The specific optimization process is as follows: Input the generated 64 positive samples, 32 spatially sampled positive samples and 96 spatially sampled negative samples into the classifier, calculate the cross-entropy loss of the binary classification, and then use the Adam optimizer to reverse The propagation algorithm updates both the classifier and the hallucination adversarial network.

H. Given a new test frame, use the region with the highest confidence of the trained classifier as the target position to complete the tracking of the current frame. The specific process is as follows: in the current test frame, use random sampling and Gaussian sampling to sample samples at the target position estimated in the previous frame. The sampled samples are input to the classifier to obtain their corresponding target confidences.

Table 1 compares the accuracy and success rate achieved by the present invention and other 9 target tracking algorithms on the OTB-2013 data set. The method of the present invention achieves excellent tracking results on mainstream datasets.

Table 1

method Accuracy (%) Success rate(%) this invention 95.1 69.6 VITAL(2018) 92.9 69.4 MCPF (2017) 91.6 67.7 CCOT (2016) 91.2 67.8 MDNet (2016) 90.9 66.8 CREST (2017) 90.8 67.3 MetaSDNet (2018) 90.5 68.4 ADNet (2017) 90.3 65.9 TRACA (2018) 89.8 65.2 HCFT (2015) 89.1 60.5

In Table 1:

VITAL corresponds to the method proposed by Y.Song et al. (Y.Song, C.Ma, X.Wu, L.Gong, L.Bao, W.Zuo, C.Shen, R.Lau, and M.-H. Yang, "VITAL: VIsual Tracking via Adversarial Learning," in Proceedings of the IEEE Conference on Computer Vision and PatternRecognition, 2018, pp.8990-8999.)

MCPF corresponds to the method proposed by T. Zhang et al. (T. Zhang, C. Xu, and M.-H. Yang, "Multi-Task Correlation Particle Filter for Robust Object Tracking," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp.4819-4827.)

CCOT corresponds to the method proposed by M. Danelljan et al. (M. Danelljan, A. Robinson, F. S. Khan, and M. Felsberg, "Beyond Correlation Filters: Learning Continuous ConvolutionOperators for Visual Tracking," in Proceedings of the European Conference on Computer Vision, 2016, pp.472-488.)

MDNet corresponds to the method proposed by H.Nam et al. (H.Nam and B.Han, "Learning Multi-domain Convolutional Neural Networks for Visual Tracking," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp.817- 825.)

CREST corresponds to the method proposed by Y. Song et al. (Y. Song, C. Ma, L. Gong, J. Zhang, R. ~ W. H. Lau, and M.-H. Yang, "CREST: Convolutional Residual Learning for Visual Tracking ,” in Proceedings of the IEEE International Conference on ComputerVision, 2017, pp.2555-2564.)

MetaSDNet corresponds to the method proposed by E.Park et al. (E.Park and A.C.Berg, "Meta-Tracker: Fast and Robust Online Adaptation for Visual Object Trackers," in Proceedings of the European Conference on Computer Vision, 2018, pp.569- 585.)

ADNet corresponds to the method proposed by S. Yun et al. (S. Yun, J. Choi, Y. Yoo, K. Yun, and J. Y. Choi, "Action-decision Networks for Visual Tracking with Deep Reinforcement Learning," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp.2711-2720.)

TRACA corresponds to the method proposed by J. Choi et al. (J. Choi, H. J. Chang, T. Fischer, S. Yun, and J. Y. Choi, "Context-aware Deep Feature Compression for High-speed Visual Tracking," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 479-488).

HCFT corresponds to the method proposed by C.Ma et al. (C.Ma, J.-B.Huang, X.Yang, and M.-H.Yang, "Hierarchical Convolutional Features for Visual Tracking," in Proceedings of the IEEE International Conference on Computer Vision, 2015, 3074-3082).

Claims (7)

1. The robust target tracking method based on hallucination confrontation network is characterized in that comprising the following steps: 1) Collect a large number of deformation sample pairs in the labeled target tracking data set as a training sample set; 2) Feature extraction is performed on all samples in the training sample set obtained in step 1) to obtain a training sample feature set; 3) using the training sample feature set, adversarial loss and the proposed deformation reconstruction loss obtained in step 2) to train the proposed hallucination adversarial network offline; The training process is as follows: first, two sets of training sample feature pairs are selected from the training sample feature set, which are expressed as

and

Learning with Illusion Adversarial Networks

and

deformation between and apply this deformation to

It is used to generate new deformed samples about the target b, and the adversarial loss is used to ensure that the generated sample distribution is similar to the distribution of the target b:

in,

_En and D _e represent the encoder and decoder parts in the proposed adversarial fantasy device, respectively; in order to make the generated samples effectively encode the deformation za , ^a deformation reconstruction loss is proposed to constrain the generated samples:

in,

Finally, the total loss function for offline training of the proposed hallucination adversarial network is:

where λ is the hyperparameter used to balance the two losses; The offline training of the hallucination confrontation network includes the following sub-steps: The parameter λ in 3.1 (formula 3) is set to 0.5; 3.2 In training, the optimizer used is Adam, the number of iterations is 5×10 ⁵ , and the learning rate is 2×10 ^-4 ; 3.3 The encoder and decoder structures of the proposed hallucination adversarial network are both three-layer perceptrons with 2048 hidden layer nodes, 9216 encoder input layer nodes, 64 encoder output layer nodes; decoder input layer nodes are 4672; the discriminant network is also a three-layer perceptron with 2048 hidden layer nodes, the number of input nodes is 9216, and the number of output nodes is 1; 4) Given the first frame annotated image in the test video, collect target samples, and use Gaussian and random sampling methods to sample positive and negative samples around the target samples; 5) Use the proposed selective deformation migration method to select the pair of samples to be migrated for the tracking target; The process of selecting the pair of samples to be migrated is: defining N _s to represent the number of video clips in the data set used to collect the pairs of deformed samples, and s _i to be the identification of the video clips, where,

Represents the number of corresponding samples in the video segment s _i ; for the feature expression ψ(s _i ) of the video segment s _i , it is calculated by the following method:

in,

is the deep feature extractor, for the target feature

Calculate the Euclidean distance between the remaining video clips representing ψ(s _i ), and select the T video clips with the nearest distance; in the selected T video clips, collect a large number of deformation sample pairs in the same way as in step 1). , forming a set D _S for subsequent target deformation migration; The selective deformation migration method includes the following sub-steps: 5.1 The deep feature extractor used when computing the feature representation of video clips

To remove the ResNet34 model of the fully connected layer; 5.2 When selecting similar video clips, the parameter T is set to 2×10 ³ ; 6) Based on the selected sample pairs to be migrated, use the offline trained hallucination confrontation network to generate deformed positive samples; 7) Use the positive and negative samples of spatial sampling and the generated positive samples to jointly train the classifier, and the resulting classification error loss is used to update the classifier and the hallucination confrontation network at the same time; 8) Given a new test frame, use the region with the highest confidence of the trained classifier as the target position to complete the tracking of the current frame. 2. the robust target tracking method based on hallucination confrontation network as claimed in claim 1, it is characterized in that in step 1) in, described in the target tracking data set with labeling, collect a large amount of deformation samples pair as training sample collection The concrete process is: : Mark the video sequence to collect a large number of target sample pairs, and a pair of samples contains the same target; in the video sequence a, first select the target sample in the t-th frame

Then randomly select the target sample in one frame in the next 20 frames as

used to form a set of deformed sample pairs

A large number of deformation sample pairs are selected to form a training sample set; the data set is the ILSVRC-2015 video target detection data set proposed by Fei-Fei Li in 2015. 3. the robust target tracking method based on hallucination confrontation network as claimed in claim 1, it is characterized in that in step 2), the step of described feature extraction is: at first the target sample is changed size to 107 using bilinear interpolation method ×107×3, and then use the neural network feature extractor φ(·) to perform feature extraction on all interpolated target samples; the structure of the feature extractor φ(·) is the VGG-M pre-trained on the Imagenet dataset The first three convolutional layers of the model. 4. the robust target tracking method based on hallucination confrontation network as claimed in claim 1 is characterized in that in step 4), the details of the sampling are: in each iteration training, the ratio of positive and negative samples is 1:3 The ratio of samples to be sampled, that is, 32 positive samples and 96 negative samples, the determination standard of positive samples is that the regional overlap rate of the sampled samples and the target samples is greater than 0.7, and the determination standard of negative samples is the regional overlap ratio of the sampled samples and the target samples. below 0.5. 5. the robust target tracking method based on hallucination confrontation network as claimed in claim 1, it is characterized in that in step 6) in, described based on the sample pair to be migrated that obtains by selection, use offline trained hallucination confrontation network to generate deformation The specific steps of positive samples are: in each iterative training, randomly select 64 pairs of samples from the set D _S , each pair of sample pairs and the target sample input against fantasy, generate corresponding deformation samples, and finally, for each iteration, a total of Generate 64 positive samples. 6. the robust target tracking method based on hallucination confrontation network as claimed in claim 1, is characterized in that in step 7) in, described using the positive and negative samples of space sampling and the positive samples generated to jointly train the classifier, its The resulting classification error loss is used to update the classifier and the hallucination adversarial network at the same time. The cross-entropy loss for classification is then used to update both the classifier and the hallucination adversarial network through a back-propagation algorithm using the Adam optimizer. 7. the robust target tracking method based on hallucination confrontation network as claimed in claim 1, is characterized in that in step 8) in, described given new test frame, use the region with the highest confidence of trained classifier as target The specific process of completing the tracking of the current frame is: in the current test frame, use random sampling and Gaussian sampling to sample samples at the target position estimated in the previous frame; the sampled samples are input into the classifier to obtain the corresponding target confidence. .

Images