激光与光电子学进展, 2019, 56 (1): 011502, 网络出版: 2019-08-01  

基于全卷积对称网络的目标尺度自适应追踪 下载: 913次

Object Scale Adaptation Tracking Based on Full-Convolutional Siamese Networks
作者单位
华北电力大学计算机系, 河北 保定 071000
摘要
针对目标追踪过程中由于目标快速运动及尺度变化导致追踪失败的问题,提出了一种基于全卷积对称网络的目标尺度自适应追踪算法。首先利用MatConvNet框架构建全卷积对称网络,使用训练好的网络得到实验图像与模板的多维特征图,两者通过互相关操作,选取置信分数最大的点为所追踪目标的中心位置;其次,对中心位置进行多尺度采样,将小于模板方差1/2的错误样本过滤掉;建立目标模板和样本概率直方图,计算模板与样本间的海林洛距离,选取合适的尺度作为目标追踪窗口的尺度。在OTB-13数据集上进行实验,与其他追踪算法性能比较,本文算法追踪成功率为0.832,精度为0.899,高于同类型深度学习追踪算法,平均追踪速度达到42.3 frame/s,满足实时性的需求;挑选包含目标快速运动或尺度变化属性的追踪序列进一步进行测试,本文算法追踪性能仍高于其他算法。
Abstract
Aim

ing at the problem of tracking failure due to fast motion and scale variation during object tracking, an object scale adaptation tracking based on full-convolutional siamese networks is proposed. First, a full-convolutional symmetric network is constructed using MatConvNet framework, and the multidimensional feature maps of template images and experimental images are obtained by using the trained networks. Through the cross-correlation operation, the point with the highest confidence score is selected as the center of the tracked target. Then, through multi-scale sampling at the center, the error samples that are less than half the template variance are filtered out. The probability histograms of target templates and samples are built. The Hellinger distance between the template and the samples is calculated, and the appropriate scale is selected as the scale of the target tracking window. Experiments on the OTB-13 dataset are carried out. Compared with other tracking algorithms, the tracking success rate of proposed method is 0.832, and the precision is 0.899, which are higher than that of other algorithms, and the average tracking speed is achieved 42.3 frame/s, meeting the needs of real-time object tracking. Selecting the tracking sequences with fast motion or scale change attributes for further testing, the tracking performance of proposed method is still higher than other algorithms.

孙晓霞, 庞春江. 基于全卷积对称网络的目标尺度自适应追踪[J]. 激光与光电子学进展, 2019, 56(1): 011502. Xiaoxia Sun, Chunjiang Pang. Object Scale Adaptation Tracking Based on Full-Convolutional Siamese Networks[J]. Laser & Optoelectronics Progress, 2019, 56(1): 011502.

引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!