基于区域候选孪生网络的红外目标跟踪算法

目标跟踪是光电设备的基本功能。为了应对跟踪任务中目标快速运动、复杂背景干扰以及遮挡的影响,不同于传统生成式方法与核相关滤波方法,本文提出一种使用深度学习的红外目标跟踪算法,使用双分支孪生特征提取网络对输入进行空间映射,经锚框划分图像区块后,分流为区域候选网络的“分类”与“回归”分支并进行特征模板匹配,对每个锚框进行分数评价后取“分类”分支中的最佳锚框,经“回归”分支进行预测边界回归后确定目标跟踪预测框,得到一种可以达到实时要求的红外单光宏观单目标跟踪算法。这种方法能够通过完全离线端到端训练整体系统参数获得,其制作过程简单,只要方法得当地进行参数精调,其性能有充分潜力可供挖掘。

Infrared target tracking algorithm based on Siamese region proposal network

Target tracking is a basic function of photoelectric equipment.In order to cope with the impact of fast target movement,complex background interference and occlusion in tracking tasks,an infrared target tracking algorithm using deep learning is proposed in this paper,which is different from traditional generative methods and kernel correlation filtering methods.The input is mapped using a double branch Siamese network into a higher dimensional space of features,and the image blocks in video frames divided by anchors are sent into the "classification" and "regression" branches of the regional proposal network.Then,correlation calculations will be conducted on "classification" branch to evaluate the matching scores between features from the template image and the search image,producing a matrix of scores for every anchor generated.The best anchor is selected after the score evaluation,and the target tracking prediction box is determined after the boundary regression from that anchor with the information of "regression" branch.An infrared single light macro single target tracking algorithm meeting the real time requirements is proposed.This approach can be obtained by training the overall system parameters end to end completely offline,is simple to produce,and has full potential for performance that can be exploited with proper parameter fine tuning of the methodology.