Target tracking algorithm based on generative adversarial network

Abstract

The invention discloses a target tracking algorithm based on a generative adversarial network, for capturing a target image required to be extracted in a video. The target tracking algorithm comprisesthe following steps: selecting and determining a target in the video; building a specific model of the target according to the characteristic values; further enhancing the precision of the characteristic values through offline training; and after determining that the target model meets the requirement by using the optimal parameter value, carrying out target search in the video, and updating thetarget by using a particle filtering method. When the target tracking algorithm is carried out, samples are directly generated through a generator in the generative adversarial network, and a probability value is generated by the discriminator to judge whether a sample is a probability value or a training value, and a particle filter framework is used to improve the tracking accuracy in a moving target tracking process, and finally an optimal parameter value of the generative adversarial network is obtained through offline training, and the output of a graph is realized by combining with the target tracking algorithm.

Description

technical field

[0001] The invention relates to the technical field of image processing, in particular to a target tracking algorithm based on a generative confrontation network. Background technique

[0002] With the rapid development of the information age, the demand for target tracking services is gradually increasing. In the field of human-computer interaction and virtual reality, contact data gloves, infrared helmets, etc. have been extensively studied, but for users to wear It is cumbersome and difficult to operate, and the tracking technology can provide a flexible, convenient and non-contact method. In the field of military guidance, the use of accurate and efficient target tracking technology can improve the accuracy of weapon strikes. In terms of sports, through multiple The high-definition camera tracks the trajectory of the ball and records it accurately, which greatly improves the viewing, fairness and accuracy of competitive sports. However, the existing movin...

Examples

Embodiment 1

[0040] Such as Figure 1-4 Shown, a target tracking algorithm based on generative adversarial networks. When performing the target tracking algorithm, the generator in the generative confrontation network directly generates samples, and the discriminator generates a probability value to determine whether the sample is a probability value or a training value. In the process of moving target tracking, the particle filter framework is used to improve tracking. Finally, through offline training, the optimal parameter value of the generative confrontation network is obtained, and combined with the target tracking algorithm to realize the output of the graph.

[0041] The target tracking algorithm is specifically divided into the following steps:

[0042] S1: Determine the target in the video by manually framing the target;

[0043] S2: Extract the characteristic value of the target through the algorithm, and build the specific model of the target according to the characteristic v...

Embodiment 2

[0047] Such as Figure 1-4 As shown, according to the steps of Embodiment 1, the extraction of target features in S2 includes global and local feature extraction. On the basis of target features, a target model is constructed. The model is divided into a generative model G and a discriminative model D. The generative model is By calculating the joint probability of the target and the sample, find the sample closest to the target model as the estimate of the current target state, and the discriminant template is to calculate the conditional probability, directly judge whether the sample is the target, and realize the target through feature extraction and target model. Effectively distinguish from background features.

[0048] The specific objective function of the generated confrontation network is as follows:

[0049]

[0050] For any G, the optimal solution of D has the following form:

[0051]

[0052] Therefore, the objective function can be simplified as:

[0053] ...

Embodiment 3

[0056] Such as Figure 1-4 As shown, according to the steps of Embodiment 1 and 2, the offline network training in S3 is carried out in a semi-supervised manner, and the best training times are obtained by setting different training times, the training times of the discriminative model D and the training times of the generated model G during training. Generative adversarial network algorithm parameter values.

technical field

[0001] The invention relates to the technical field of image processing, in particular to a target tracking algorithm based on a generative confrontation network. Background technique

[0002] With the rapid development of the information age, the demand for target tracking services is gradually increasing. In the field of human-computer interaction and virtual reality, contact data gloves, infrared helmets, etc. have been extensively studied, but for users to wear It is cumbersome and difficult to operate, and the tracking technology can provide a flexible, convenient and non-contact method. In the field of military guidance, the use of accurate and efficient target tracking technology can improve the accuracy of weapon strikes. In terms of sports, through multiple The high-definition camera tracks the trajectory of the ball and records it accurately, which greatly improves the viewing, fairness and accuracy of competitive sports. However, the existing movin...

Examples

Embodiment 1

[0040] Such as Figure 1-4 Shown, a target tracking algorithm based on generative adversarial networks. When performing the target tracking algorithm, the generator in the generative confrontation network directly generates samples, and the discriminator generates a probability value to determine whether the sample is a probability value or a training value. In the process of moving target tracking, the particle filter framework is used to improve tracking. Finally, through offline training, the optimal parameter value of the generative confrontation network is obtained, and combined with the target tracking algorithm to realize the output of the graph.

[0041] The target tracking algorithm is specifically divided into the following steps:

[0042] S1: Determine the target in the video by manually framing the target;

[0043] S2: Extract the characteristic value of the target through the algorithm, and build the specific model of the target according to the characteristic v...

Embodiment 2

[0047] Such as Figure 1-4 As shown, according to the steps of Embodiment 1, the extraction of target features in S2 includes global and local feature extraction. On the basis of target features, a target model is constructed. The model is divided into a generative model G and a discriminative model D. The generative model is By calculating the joint probability of the target and the sample, find the sample closest to the target model as the estimate of the current target state, and the discriminant template is to calculate the conditional probability, directly judge whether the sample is the target, and realize the target through feature extraction and target model. Effectively distinguish from background features.

[0048] The specific objective function of the generated confrontation network is as follows:

[0049]

[0050] For any G, the optimal solution of D has the following form:

[0051]

[0052] Therefore, the objective function can be simplified as:

[0053] ...

Embodiment 3

[0056] Such as Figure 1-4 As shown, according to the steps of Embodiment 1 and 2, the offline network training in S3 is carried out in a semi-supervised manner, and the best training times are obtained by setting different training times, the training times of the discriminative model D and the training times of the generated model G during training. Generative adversarial network algorithm parameter values.

Images