激光与光电子学进展, 2019, 56 (22): 221502, 网络出版: 2019-11-02  

基于时序一致和空间剪裁的多特征相关滤波跟踪算法 下载: 873次

Tracking Algorithm of Correlation Filter with Multiple Features Based on Temporal Consistency and Spatial Pruning
作者单位
江南大学物联网工程学院模式识别与计算智能国际联合实验室, 江苏 无锡 214122
摘要
为提升相关滤波跟踪算法在目标遮挡、背景嘈杂及目标形变等干扰下的跟踪精度,提出一种基于时序一致和空间剪裁的多特征相关滤波跟踪算法。在训练阶段利用二值矩阵掩模对滤波器模板的能量分布进行裁剪,使模板信息更加集中于目标区域,从而缓解循环样本造成的边界效应;利用l2范数作为时序一致模型对相邻帧的滤波器建立平滑性约束,使滤波器模板学习到相邻帧目标的上下文信息,增加算法的抗干扰能力;为进一步提升目标模板的表达能力,将包含丰富语义信息的ResNet50深度特征引入到跟踪框架中,通过主成分分析法对提取到的深度特征进行降维,采用传统特征结合深度特征的方式提升跟踪结果的精确度和稳健性。将本文算法与5种算法进行对比实验,验证了本文算法在处理目标遮挡、背景嘈杂及目标形变等干扰时的稳健性。
Abstract
Aim

ing to improve the tracking accuracy of the correlation filter tracking algorithm when faced with occlusion, background clutter, and deformation of the object target, this study proposes a correlation filter tracking algorithm for multiple features based on temporal consistency and spatial pruning. First, in the training stage, the energy distribution of the filter template is pruned using the binary matrix mask to make the template information more concentrated in the target area, which alleviates the boundary effect caused by the cyclic shifted samples. Second, the l2-norm is used as the temporal consistency model to establish smoothness constraints for the filters of two consecutive frames so that filter templates can learn the context information of consecutive-frame target and increase the anti-interference ability of the algorithm. To further improve the expressive ability of the target template, ResNet50 deep features, which contain rich semantic information, are introduced into the tracking framework. Principal component analysis is used to reduce the dimension of the extracted depth feature, and traditional features in combination with deep features improve the accuracy and robustness of the tracking results. A comparison of the proposed algorithm with five existing algorithms verifies the proposed tracking algorithm’s robustness in dealing with distractors such as target occlusion, background clutter, and deformation.

王译萱, 吴小俊, 徐天阳. 基于时序一致和空间剪裁的多特征相关滤波跟踪算法[J]. 激光与光电子学进展, 2019, 56(22): 221502. Yixuan Wang, Xiaojun Wu, Tianyang Xu. Tracking Algorithm of Correlation Filter with Multiple Features Based on Temporal Consistency and Spatial Pruning[J]. Laser & Optoelectronics Progress, 2019, 56(22): 221502.

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!