Bimodal target tracking algorithm based on feature level and decision level fusion

Abstract

The invention relates to a bimodal target tracking algorithm based on feature level and decision level fusion. The bimodal target tracking algorithm comprises the following steps: constructing a SiamDL two-level fusion attention network structure; acquiring a template image; acquiring a search area image; extracting image depth features; performing interaction on the depth features of the multiple domains; performing classification constraint on the interacted features; modulating a classification result; fusing features; modulating the fusion features; and performing classifying and regression. By introducing a double-layer fusion attention mechanism, a double-stage balance module is provided, and the weight ratio of two modes can be more reasonably balanced by using decision-level and feature-level information; a cross-domain twinborn attention mechanism is introduced, a multi-domain sensing module is provided, template features can be adaptively updated, rich context information of a mode domain and a time domain is utilized, the feature representation capacity of a network is improved, high-speed operation and an excellent tracking result are achieved, and the capacity of a tracker for coping with complex scenes is improved.

Description

technical field [0001] The invention relates to a dual-mode target tracking algorithm based on feature-level and decision-level fusion, and belongs to the technical field of target tracking. Background technique [0002] Object tracking Given an initial object template, estimating its position and size in subsequent frames is an important task in the field of computer vision. With the advent of correlation filtering and deep learning, visible light object tracking has achieved great development. However, when the visible light modal characteristics are not enough to reveal the target information, such as dark light, exposure or submerged in the background, the visible light tracking effect will be greatly reduced. [0003] Most of the time, the infrared mode is rich in the structural information of the target, and the visible light mode is rich in the structure and texture information of the target. Adding infrared modal information For a tracker, visible light can supplem...

AI Technical Summary

Problems solved by technology

[0005] Many works directly use feature-level fusion strategies to calculate the channel weight ratio of fused features, which inevitably contains a lot of background information, which greatly affects the calculation of this weight ratio.

Images