Real-time video target tracking algorithm based on multilayer attention mechanism

Abstract

The invention provides a real-time video target tracking algorithm based on a multilayer attention mechanism. The real-time video target tracking algorithm comprises the following steps of inputting aseries of template images and search images; constructing a multilayer attention convolutional neural network as a matching function; calculating the similarity between the template image and the search image by utilizing related operations; using a logic loss function as an optimized objective function; optimizing the network parameters by adopting momentum random gradient descent; fixing the model parameters during the online tracking of the convolutional network, taking a first frame of target as a template to be inputted into the network, inputting the subsequent frames sequentially intothe network to carry out the correlation calculation with the output obtained by the first frame, and obtaining a response score graph; and finding the position with the highest score in the responsegraph as the position of the target, and amplifying the target to the size the same as that of the original graph to obtain the tracking target. According to the multilayer attention twinning networktracking algorithm provided by the invention, an attention mechanism can be effectively utilized to separate a target from a background, and the accuracy and the generalization of the tracker are remarkably improved under the conditions of target rotation, deformation and partial shielding.

Description

technical field [0001] The invention relates to a real-time video target tracking algorithm based on a multi-layer attention mechanism, which belongs to the field of image processing. Background technique [0002] The goal of object tracking is to track any temporally changing object specified with a bounding box only in the first frame. Object tracking is one of the important components of various applications in the field of computer vision, such as unmanned driving, video surveillance, human-computer interaction and other fields. Although research in this field has made significant progress in the past decade, due to the arbitrary potential changes of the target and the unknown and continuously changing background of the video, partial occlusion, fast motion, Due to the interference of factors such as scale changes, the tracking of arbitrary targets has been very challenging so far. [0003] In recent years, as deep learning has achieved great success in the field of co...

AI Technical Summary

Problems solved by technology

However, the target tracking algorithm (SiamFC) method based on the fully convolutional twin network uses a shallower convolutional neural network designed for classification tasks during offline training, which makes the representation ability of the model weak, and serious deformation occurs in the target. Or when similar objects appear, the model is prone to drift due to the weak discriminative power of the model, resulting in tracking failure

Images