A computer vision application-oriented anchor-frame-free target tracking algorithm

Abstract

The invention mainly provides a computer vision application-oriented anchor-frame-free target tracking algorithm, and the method based on a twin architecture can obtain more remarkable performance while keeping a good speed, so that wide attention in the visual tracking field is aroused. However, twin network branches are generally independent and lack information interaction, which limits further improvement of model performance. In order to enhance the cooperation capability of twinborn network branches, the invention provides a cross-aware network model based on a twinborn architecture, and the cross-aware network model is called as SiamCrossObject-Aware Network. In SiamCross, on the basis of mutual supervision of twinning network branches, a brand-new twinning cross-aware sub-network (Siamese Crosss-Aware Network, SCAN) is designed. The SCAN can enable the features extracted by the twin network to be effectively perceived and distinguish the foreground and the background. A tracking task is directly defined as classification and regression of each pixel based on an anchor-frame-free region candidate network, and network branch features can focus on local and global spatial information of a target.

Description

technical field [0001] The invention relates to the field of target tracking, in particular to an anchor-free target tracking algorithm for computer vision applications. Background technique [0002] Visual object tracking is a fundamental task in computer vision. Given an arbitrary object state in the initial frame of the sequence, the tracker needs to accurately estimate the object state in subsequent frames. Object tracking has a wide range of applications in areas such as automatic surveillance, vehicle navigation, robot sensing, human-computer interaction, and augmented reality. Although great progress has been made in visual tracking, robust visual tracking is still very challenging due to the influence of many factors such as illumination changes, occlusion and background interference. [0003] The Siamese network-based tracking algorithm regards the visual target tracking problem as learning a general similarity score map, because it has achieved a good balance bet...

AI Technical Summary

Problems solved by technology

Some current research attempts to improve the performance of the tracker by combining features between branches, but did not fully exploit the potential of the Siamese-RPN architecture.

Images