*2.2. The Triple Structure Siamese Network for Velocity Spectra Lateral Target Tracking*

For lateral target tracking in velocity spectra, the prior materials are the images of the energy clusters in the given windows. The tracking algorithm needs to overcome the challenges of target deformation, background interference, scale change, and angle rotation, and also needs to take into account the accuracy and efficiency. The Siamese network can find a set of parameters, so that the similarity of input images is large when they belong to the same category and small when they belong to different categories. Another advantage of the Siamese network is that the input is a pair of images rather than an image, so it can naturally increase the amount of training data and make full use of limited datasets to train, which is very important in the field of target tracking. In this paper, we improved the traditional Siamese network structure by adding an input update branch. A triple structure Siamese network for velocity spectra lateral target tracking is presented. We set the target area as a positive sample, and the background area and other morphological energy clusters as negative samples. Firstly, the tracking algorithm extracts the features of the target through a series of convolution operations and trains the classifier. Next, the well-trained classifier is used to find the most similar region in the velocity spectra of different CDP positions. Finally, the lateral extrapolation path of inversion parameters can be obtained through the change trend of the target positions. The added update branch can take the prediction result of the current position as an input and update the initial target according to the current tracking result, so as to adapt to the lateral change of the velocity spectrum.
