加权间隔结构化支持向量机目标跟踪算法

发布时间：2018-03-20 11:15

本文选题：目标跟踪　切入点：结构化支持向量机　出处：《中国图象图形学报》2017年09期 　论文类型：期刊论文

【摘要】：目的目标在跟踪过程中,各种因素的干扰使得跟踪结果存在不确定性。因此,将跟踪过程中所提取样本的可靠性融入跟踪模型中,有助于克服低可靠性样本对跟踪算法的影响。为此,基于最近的结构化支持向量机(SSVM)跟踪算法,提出一种包含样本置信度的加权间隔结构化支持向量机跟踪模型(WMSSVM),以增强SSVM跟踪算法性能。方法首先,基于打分和位置重合率估计样本可靠性;其次,建立WMSSVM模型处理具有不同置信度的跟踪样本训练问题,并采用对偶坐标下降优化算法求解跟踪模型。结果在包含100个视频的OTB100跟踪数据集上进行测试,提出的WMSSVM跟踪器与基准跟踪器Scale-DLSSVM相比,在精准度和成功率两个指标上分别提高了1%和2%。与最近的跟踪算法相比,提出的方法也表现出更好的性能。结论本文首次将样本的可靠性融入结构化支持向量机跟踪模型,并提出一种基于加权间隔的结构化支持向量机跟踪模型及其优化求解方法,在包含100个视频序列的跟踪数据集上验证了提出方法的有效性,本文提出的算法能够适应复杂场景下的跟踪任务,并在背景混杂、目标形变、遮挡、运动模糊、目标出界、快速位移等类别的视频中表现出优异的性能。
[Abstract]:Objective in the process of target tracking, the disturbance of various factors makes the tracking results uncertain. Therefore, the reliability of the samples extracted in the tracking process is incorporated into the tracking model. It is helpful to overcome the influence of low reliability samples on tracking algorithm. To this end, based on the most recent structured support vector machine (SSVM) tracking algorithm, A weighted interval structured support vector machine tracking model including sample confidence is proposed to enhance the performance of SSVM tracking algorithm. Firstly, the reliability of the sample is estimated based on scoring and position coincidence rate. The WMSSVM model is established to deal with the training problem of tracking samples with different confidence levels, and the dual coordinate descent optimization algorithm is used to solve the tracking model. The results are tested on the OTB100 tracking data set containing 100 videos. Compared with the reference tracker (Scale-DLSSVM), the proposed WMSSVM tracker improves the accuracy and success rate by 1% and 2, respectively. The proposed method also shows better performance. Conclusion in this paper, the reliability of samples is first integrated into the structured support vector machine tracking model, and a structured support vector machine tracking model based on weighted interval and its optimization solution are proposed. The effectiveness of the proposed method is verified on the tracking data set of 100 video sequences. The proposed algorithm can adapt to the tracking tasks in complex scenes, and can be used in background clutter, object deformation, occlusion, motion blur, target out of bounds. Fast displacement and other categories of video show excellent performance.
【作者单位】：西北农林科技大学信息工程学院;西安电子科技大学综合业务网理论及关键技术国家重点实验室;
【基金】：西安电子科技大学综合业务网理论及关键技术国家重点实验室开放基金项目(ISN17-08)
【分类号】：TP18;TP391.41

【相似文献】