基于L-GEM径向基函数神经网络的目标检测跟踪算法研究

发布时间：2018-01-06 00:02

本文关键词：基于L-GEM径向基函数神经网络的目标检测跟踪算法研究　出处：《浙江大学》2016年硕士论文　论文类型：学位论文

【摘要】：运动目标检测跟踪在智能视频监控等诸多领域起到十分重要的作用。在过去的几年中,尽管大量的检测跟踪相关算法被提出,但在算法的实际应用过程中会由于场景情况多样、算法计算量大等因素难以满足实时检测跟踪的需求,这也是目前为止,智能视频监控等没有取得广泛应用的原因。面对这样的情况,如何设计一款有效准确的实时目标检测跟踪算法有着重大的研究意义和实用意义。针对目标检测跟踪的实时性以及准确性,本文提出了一种基于当前两种主流框架相结合的算法来完成目标的检测跟踪,并在传统的目标运动区域提取与目标检测两个模块中提出了以下几种改进算法：一、使用一种改进的混合高斯背景模型来完成运动区域的提取。考虑到混合高斯背景模型在面对背景突变等情况下建模效果会变差,本文对算法进行了改进,即通过调整背景突变时模型的学习速率来削弱模型对光照变化的敏感度使模型具有更好的适应性。二、使用基于L-GEM的RBFNN分类器来完成运动目标的训练检测工作。基于L-GEM的RBFNN分类器因其良好的泛化能力和极快的学习收敛速度被广泛用于诸多应用领域,但对目标检测跟踪研究算法分析之后发现至今还没有在这方面取得应用。三、使用基于随机敏感度的重抽样来平衡训练过程中目标背景样本不平衡问题。对于帧图像来说,样本内的目标区域往往远小于背景区域,而训练过程中的误分将产生极大的代价,目前的很多分类器如RBFNN、SVM等都没有考虑到这个问题。本文使用基于随机敏感度的重抽样来平衡目标背景训练过程中样本不平衡问题,进而提高训练器的分类性能。本文从以上三个方面对当前的运动目标检测跟踪算法做了改进,实验结果也证明本文算法在保证检测跟踪精度的同时,也保证了算法的实时性。
[Abstract]:Moving target detection and tracking plays an important role in many fields, such as intelligent video surveillance. In the past few years, although a large number of detection and tracking algorithms have been proposed. However, in the practical application of the algorithm, it is difficult to meet the needs of real-time detection and tracking because of the diversity of the scene and the large amount of calculation of the algorithm, which is also up to now. Intelligent video surveillance and other reasons are not widely used. Faced with such a situation. How to design an effective and accurate real-time target detection and tracking algorithm is of great significance and practical significance, aiming at the real-time and accuracy of target detection and tracking. In this paper, an algorithm based on the combination of two main frameworks is proposed to complete target detection and tracking. And in the traditional target moving region extraction and target detection two modules proposed the following improved algorithms: 1. An improved mixed Gao Si background model is used to extract the moving region. Considering that the mixed Gao Si background model will become worse in the face of background mutation, the algorithm is improved in this paper. That is, by adjusting the learning rate of the model with background mutation to weaken the sensitivity of the model to light change, the model has better adaptability. The RBFNN classifier based on L-GEM is used to complete the training and detection of moving target. The RBFNN classifier based on L-GEM has been widely used because of its good generalization ability and fast learning convergence speed. It is widely used in many application fields. But after analyzing the algorithm of target detection and tracking, it is found that it has not been applied in this field. 3. Resampling based on random sensitivity is used to balance the imbalance of target background samples in the process of training. For frame images, the target region in the sample is often much smaller than the background region. However, the misclassification in the training process will have a great cost. Many current classifiers such as RBFNN. SVM et al. did not consider this problem. In this paper, we use resampling based on random sensitivity to balance the sample imbalance in the process of target background training. And then improve the classification performance of the trainer. This paper improves the current moving target detection and tracking algorithm from the above three aspects, the experimental results also prove that the algorithm can ensure the accuracy of detection and tracking at the same time. It also ensures the real-time performance of the algorithm.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP183

【相似文献】