基于似物性采样和深度学习的目标跟踪算法研究

发布时间：2018-05-15 16:58

本文选题：目标跟踪 + 卷积神经网络　；参考：《华侨大学》2017年硕士论文

【摘要】：目标跟踪是计算机视觉研究的一个重要方向,由于具有广泛的应用前景,如智能视频监控,人机交互,机器人等,吸引了国内外大量学者研究的热潮,但由于目标跟踪面临各种复杂因素,如何设计一个鲁棒的目标跟踪算法还需要不断地努力。近几年,由于深度学习(Deep Learning)的发展,基于深度学习的目标跟踪算法取得了很大的突破,但由于在目标跟踪中获取训练样本较少,需要大量数据离线训练深度模型,既耗时,且离线学到的特征表示不一定对在线的目标跟踪有益。同时由于实际数据往往存在噪声,基于深度模型学到的特征表示有可能存在与任务不相关的模式,影响算法的性能。针对上述提到的问题,本文提出了两种新颖的目标跟踪算法,本文研究的主要内容如下:1)提出了基于似物性采样和PCA卷积神经网络的目标跟踪算法。近几年基于深度学习(Deep Learning)的目标跟踪算法受到广泛关注,然而缺少训练样本成为最大的瓶颈,通过离线训练深度模型既耗时,且学到的基本特征表示不一定对在线目标跟踪有益。针对这个问题,本文提出了一种基于PCA卷积神经网络的目标表观模型,该网络不需要离线训练,卷积模板通过主成成分分析(PCA)算法在线学习得到,同时对学到的特征采用软阈值方法(soft-thresholding)获得更加鲁棒的稀疏特征表示。为了进一步提升跟踪算法性能,基于物体边缘轮廓信息的似物性采样方法计算目标物体的似物性分值,最后通过多线索融合机制估计出目标的最佳状态,并在常用的目标跟踪数据库(OTB-50)上做了实验分析,验证了算法的有效性。2)提出了基于似物性采样和阀门卷积深度网络的目标跟踪算法。目标跟踪算法的性能主要取决于物体表观的特征表示(包括特征学习和特征选择)。跟踪算法的一个通常假设是原始视频数据是没有噪声的,然而这个假设太过严格,尤其是在复杂的现实场景中,如果一开始就没有好的原始特征数据,有可能会降低算法学习的能力。针对这个问题,本文设计了一种新颖的基于阀门卷积深度网络的目标表观模型,该模型通过阀门机制能够选择与任务相关模式的特征,并且能够将特征学习和特征选择统一于同一框架。同时为了减缓目标发生“漂移”,融合了基于物体边缘轮廓信息的似物性采样方法来进一步提升算法的准确率,并在常用的目标跟踪数据库(OTB-50)上做了实验分析,验证了算法的有效性。
[Abstract]:Target tracking is an important direction of computer vision research. Because of its wide application prospect, such as intelligent video surveillance, human-computer interaction, robot, etc., it attracts a lot of scholars at home and abroad. However, due to the complexity of target tracking, it is necessary to design a robust target tracking algorithm. In recent years, due to the development of Deep Learning (Deep Learning), the target tracking algorithm based on deep learning has made a great breakthrough. However, because of the small number of training samples obtained in target tracking, it requires a lot of data to train depth model offline, which is time-consuming. And the offline feature representation is not necessarily useful for online target tracking. At the same time, because of the noise in the actual data, the feature representation based on the depth model is likely to have a pattern that is not related to the task, which affects the performance of the algorithm. Aiming at the above mentioned problems, two novel target tracking algorithms are proposed in this paper. The main contents of this paper are as follows: 1) A target tracking algorithm based on quasi-physical sampling and PCA convolution neural network is proposed. In recent years, the target tracking algorithm based on Deep Learning Learning has received wide attention. However, the lack of training samples has become the biggest bottleneck, and it is time-consuming to use off-line training depth model. And the basic feature representation is not necessarily beneficial to online target tracking. In order to solve this problem, this paper presents a target representation model based on PCA convolution neural network, which does not require offline training, and the convolution template is obtained by the principal component analysis (PCA) algorithm. At the same time, soft threshold method is used to obtain a more robust sparse feature representation. In order to further improve the performance of the tracking algorithm, the physical property sampling method based on the contour information of the object edge is used to calculate the object-like property score of the target object. Finally, the best state of the target is estimated by the multi-clue fusion mechanism. Based on OTB-50), a target tracking algorithm based on quasi-physical sampling and valve convolution depth network is proposed, which validates the validity of the algorithm. The performance of the target tracking algorithm mainly depends on the feature representation (including feature learning and feature selection). One of the common assumptions of the tracking algorithm is that the original video data is noise-free, but this assumption is too strict, especially in complex real-world scenarios, if there is no good original feature data at first. It is possible to reduce the ability of algorithm learning. To solve this problem, a novel target model based on valve convolution depth network is designed in this paper. The model can select the characteristics of task-related modes through valve mechanism. And feature learning and feature selection can be unified in the same framework. At the same time, in order to reduce the "drift" of the target, this paper combines the object-like sampling method based on the contour information of the object edge to further improve the accuracy of the algorithm, and makes an experimental analysis on the commonly used target tracking database (OTB-50). The validity of the algorithm is verified.
【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】