| 引用本文: |
苑红晓1，冯玉芳1，潘峰1，殷宏2，白景波2，曾祥熙1.基于轻量卷积神经网络的目标跟踪改进算法[J].陆军工程大学,2023,(1):77-85
[点击复制] |
|
| |
|
|
| 本文已被：浏览 188次 下载 242次 |
| 基于轻量卷积神经网络的目标跟踪改进算法 |
| 苑红晓1，冯玉芳1，潘峰1，殷宏2，白景波2，曾祥熙1 |
| (1.32125部队，山东 济南 250004;2.陆军工程大学 指挥控制工程学院，江苏 南京 210007) |
| 摘要:针对现有目标跟踪算法在跟踪过程中遇到目标形变、遮挡等干扰属性导致不能对目标进行有效跟踪的问题，提出一种基于轻量卷积神经网络（lightweight convolutional neural network,LWCN）的目标跟踪改进算法。首先利用改进的卷积神经网络对模板图片和跟踪图片进行特征提取，并将不同层次的特征图充分利用，解决了随着网络加深而导致部分特征丢失问题；其次融合CN特征和HOG特征作为相关滤波器中目标特征表达，增强在不同干扰属性下的目标描述能力；再次通过最大响应值对当前目标位置和目标尺度进行判断，并决定是否更新滤波器模板；最后将LWCN算法与其他算法在OTB50、OTB100、UAV123等数据集上进行性能对比实验。实验结果表明，LWCN算法具有较好的稳定性和实时性，并在遇到形变、遮挡、光线和背景变化时，跟踪结果优于大部分算法。 |
| 关键词: 目标跟踪 轻量CNN 特征提取 特征融合 |
| DOI：10.12018/j .issn.2097-0730.20220314001 |
| 投稿时间：2022-03-14 |
| 基金项目:国家重点研发计划（2017YFB0802800) |
|
| Object Tracking Algorithm Based on Lightweight Convolutional Neural Network |
| YUAN Hongxiao1，FENG Yufang1，PAN Feng1，YIN Hong2，BAI Jingbo2，ZENG Xiangxi1 |
| (1.Unit 32125 of PLA，Jinan 250004，China;2.College of Command & Control Engineering，
Army Engineering University of PLA，Nanjing 210007，China) |
| Abstract: To solve the problem that existing object tracking algorithms cannot effectively function during tracking when encountering some interference properties such as object deformation and occlusion, this paper proposes an improved object tracking algorithm based on lightweight convolutional neural network (LWCN). Firstly, the improved convolutional neural network was used to extract features from the template images and tracking images. To solve the problem that some features will be lost with network deepening, this network fused the object features from different levels to make full use of these features. Secondly, to enhance the ability to describe the object in complex situations, the fusion of CN and HOG features was used as the object feature expression in the correlation filter. Thirdly, the maximum response value was used to judge the current object position and size, and determine whether the filtering template needs to be updated. Lastly, the LWCN was compared with other algorithms on OTB50, OTB100 and UAV123 test datasets. The results show that the LWCN has better stableness and real-time performance and it functions better than most algorithms in conditions of object deformation, occlusion, light and background changes. |
| Key words: object tracking lightweight CNN feature extraction feature fusion |
|
