基于卷积网络的多模板鲁棒目标跟踪方法研究

发布时间：2018-03-01 04:47

本文关键词： 目标跟踪模型更新卷积网络归一化加权多模板　出处：《昆明理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：目标跟踪是计算机视觉研究领域的热点之一,本文首先介绍了目标跟踪技术的研究背景和发展现状,描述了目标跟踪中特征提取和运动估计两个关键部分,并简要分析了跟踪过程可能遇到的难点。大部分目标跟踪算法可分为两个步骤:目标特征提取和运用目标特征实现跟踪的算法。在复杂背景或目标被局部遮挡时,目标特征信息的唯一性和稳定性下降,使得跟踪算法不能准确的区分目标和背景,导致跟踪算法失效;目标模型更新策略同样是维系稳定跟踪的重点,跟踪算法中不可缺少的一部分。目前的模型更新方法仅依赖来于上一帧或最近帧定位到的目标信息,跟踪的历史信息未充分利用,当发生遮挡和形变后,跟踪算法不能准确的重新定位目标。针对上述问题本文提出算法采用了归一化距离加权函数和多模板模型更新策略,归一化加权方法通过目标模板像素到模板中心的距离构建加权函数,在复杂背景时可对目标特征进行增强,减少背景对目标特征信息的干扰;多模板模型更新策略可以在跟踪过程提供更完备的目标模型匹配信息,本文基于该更新模型结合卷积网络提出一种新的运动目标跟踪方法。与目前热点运动目标跟踪方法在V0T2015运动目标跟踪测试视频集下的对比实验表明,本文方法对于遮挡现象和目标自身形变具有较强的鲁棒性和较高的准确性。
[Abstract]:Target tracking is one of the hotspots in the field of computer vision. Firstly, this paper introduces the background and development of target tracking technology, and describes two key parts of feature extraction and motion estimation in target tracking. Most of the target tracking algorithms can be divided into two steps: target feature extraction and target feature tracking algorithm. The uniqueness and stability of the target feature information is decreased, which makes the tracking algorithm unable to distinguish between the target and the background accurately, which leads to the failure of the tracking algorithm. The updating strategy of the target model is also the key to maintain the stable tracking. The current model updating method only depends on the target information located in the previous or most recent frame, and the historical information of the tracking is not fully utilized, when occlusion and deformation occur, The tracking algorithm can not accurately relocate the target. In view of the above problems, this paper proposes a normalized distance weighting function and a multi-template model updating strategy. The normalized weighting method constructs the weighting function through the distance between the target template pixel and the template center, which can enhance the target feature and reduce the interference of the background to the target feature information when the background is complex. Multi-template model updating strategy can provide more complete target model matching information in the tracking process. Based on the updated model and convolutional network, a new moving target tracking method is proposed in this paper, which is compared with the current hot moving target tracking method in V0T2015 moving target tracking video set. The proposed method is robust and accurate to occlusion and target deformation.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】