短时域目标识别与快速跟踪神经网络研究  

Target recognition and fast tracking neural networks under short term

作　　者：曹昭睿 张慧 张伟 郝永平[1] 胡晓阳[1] 王俊杰[3] CAO Zhaorui;ZHANG Hui;ZHANG Wei;HAO Yongping;HU Xiaoyang;WANG Junjie(College of Equipment Engineering,Shenyang Ligong University,Shenyang 110158,China;College of Mechanical Engineering,Shenyang Ligong University,Shenyang 110158,China;College of Science,Shenyang Ligong University,Shenyang 110158,China)

机构地区：[1]沈阳理工大学装备工程学院,沈阳110158 [2]沈阳理工大学机械工程学院,沈阳110158 [3]沈阳理工大学理学院,沈阳110158

出　　处：《兵器装备工程学报》2025年第3期182-188,共7页Journal of Ordnance Equipment Engineering

基　　金：辽宁省教育厅基本科研项目(JYTMS20230202,LJ212410144050);沈阳理工大学引进高层次人才科研支持计划项目(1010147001254)。

摘　　要：针对多模态检测任务下全尺寸目标识别与跟踪网络串联结构较大、整合程度差、计算效率低的问题,克服跟踪算法对目标运动持续时间与轨迹积累的依赖,面向无人机、智能巡飞弹等无人装备对多模态网络一体化、轻量化与快速相应的需求,提出了一种基于短时域的目标识别与快速跟踪神经网络。以相邻两帧图像作为输入,融合目标语义信息与运动趋势信息间的关联特征;通过可共享特征提取网络,降低多模态目标检测网络的结构复杂性;采用含有静态与动态链路的双分支推理网络,同步完成当前帧内目标识别计算与未来帧内目标位置预测。实验结果表明:所提出的算法识别准确率可达到95.4%,位置预测准确率可达到90.9%,能够在低算力支持下赋予智能武器目标高效识别与快速跟踪计算能力。Aiming at the problems of large series structure,low integration and computational efficiency of the full-scale target recognition and tracking network under the multi-modal detection task,overcoming the dependence of the tracking algorithm on the target motion duration and trajectory accumulation,and facing the requirements of unmanned aerial vehicles,intelligent cruise missiles and other unmanned equipment for multi-modal network integration,lightweight and fast response,a neural network for object recognition and fast tracking under short term is proposed.Using adjacent two frames of images as inputs,the correlation features between target semantic information and motion trend information are fused.By the design of sharable feature extraction network,the structural complexity of multi-modal object detection networks is reduced.Using a dual branch inference network with static and dynamic links,the target recognition in current and position prediction in future are completed simultaneously.The experimental results show that the proposed algorithm achieves an accuracy rate of 95.4%for object recognition and 90.9%for location prediction,and capable of endowing intelligent weapons with efficient target recognition and fast tracking computing capabilities with low computing power support.

关 键 词：目标识别 目标跟踪 位置预测 深度学习 机器视觉 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]