基于压缩感知的视频目标跟踪研究

发布时间：2018-02-10 23:06

本文关键词： 压缩跟踪特征在线选择协方差矩阵粒子滤波样本加权　出处：《中国民航大学》2017年硕士论文　论文类型：学位论文

【摘要】：目标跟踪作为计算机视觉领域的研究热点与难点已在智能人机交互、视觉导航、智能视频监控等领域得到了广泛的应用。但是,研制出一种在各种复杂场景下均能够实现稳定、快速、高精度跟踪的目标跟踪算法仍然是一项极具挑战性的课题。近年来,压缩跟踪算法因其良好的跟踪性能成为了研究热点。为了提高复杂场景下压缩跟踪算法的跟踪性能,对其进行了研究,主要成果如下:为了提高压缩跟踪算法在光照变化、遮挡场景下的跟踪稳定性和准确性,提出一种结合特征在线选择与协方差矩阵的压缩跟踪算法。首先,在特征提取阶段引入基于Hellinger距离的特征在线选择方法,在特征池中动态选择那些置信水平较高的特征用于构建朴素贝叶斯分类器。然后,在压缩跟踪的框架下融合协方差矩阵以增强算法对目标的表达能力,把Haar-like特征和协方差矩阵相结合构建目标模型,取最大响应值所对应的候选样本作为跟踪结果。最后,优化分类器参数的更新方式,根据目标模板与跟踪结果的相似度来自适应地更新分类器参数。针对压缩跟踪算法无法适应目标尺度的变化以及没有考虑样本权重的问题,提出一种基于粒子滤波与样本加权的压缩跟踪算法。首先,对原始压缩跟踪算法中的压缩特征进行改进,提取归一化矩形特征用于构建目标表观模型。然后,引入样本加权的思想,根据正样本与目标之间距离的不同赋予正样本不同的权重,提高分类器的分类精度。最后,在粒子滤波的框架下融合尺度不变压缩特征进行动态状态估计,在粒子预测阶段利用一个二阶自回归模型对粒子状态进行估计与预测,借助观测模型对粒子状态进行更新,并且对粒子进行重采样以防止粒子退化。实验结果表明,相比于原始压缩跟踪算法,结合特征在线选择与协方差矩阵的压缩跟踪算法在光照变化和遮挡场景下具有更高的跟踪鲁棒性和准确性;基于粒子滤波与样本加权的压缩跟踪算法能够更好的跟踪目标尺度的变化,提高了的跟踪稳定性和准确性。
[Abstract]:As a hot and difficult point in the field of computer vision, target tracking has been widely used in the fields of intelligent human-computer interaction, visual navigation, intelligent video surveillance and so on. Fast and high precision target tracking algorithm is still a challenging subject. In recent years, compression tracking algorithm has become a research hotspot for its good tracking performance. The main results are as follows: in order to improve the tracking stability and accuracy of the compression tracking algorithm under the illumination variation and occlusion scene, a compression tracking algorithm combining online feature selection and covariance matrix is proposed. In the stage of feature extraction, an online feature selection method based on Hellinger distance is introduced, and those features with high confidence level are dynamically selected in the feature pool to construct naive Bayes classifier. In the framework of compressed tracking, the covariance matrix is fused to enhance the ability of the algorithm to express the target. The target model is constructed by combining the Haar-like features with the covariance matrix, and the candidate samples corresponding to the maximum response value are taken as the tracking results. According to the similarity between the target template and the tracking result, the classifier parameters are updated adaptively. The compression tracking algorithm can not adapt to the change of the target scale and does not consider the weight of the sample. A compression tracking algorithm based on particle filter and sample weighting is proposed. Firstly, the compression feature of the original compression tracking algorithm is improved, and the normalized rectangular feature is extracted to construct the target apparent model. The idea of weighted samples is introduced to give different weights to positive samples according to the distance between positive samples and targets. Finally, the classification accuracy of the classifier is improved. In the framework of particle filter, the dynamic state estimation is carried out by integrating the scale-invariant compression features. In the phase of particle prediction, a second-order autoregressive model is used to estimate and predict the particle state, and the observation model is used to update the particle state. The experimental results show that, compared with the original compression tracking algorithm, the proposed algorithm is better than the original compression tracking algorithm. The compression tracking algorithm based on feature online selection and covariance matrix has higher tracking robustness and accuracy under varying illumination and shading scenes. The compressed tracking algorithm based on particle filter and sample weighting can better track the change of target scale and improve the tracking stability and accuracy.
【学位授予单位】：中国民航大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】