基于改进的TLD目标跟踪算法研究

发布时间：2018-03-13 23:11

本文选题：跟踪-学习-检测　切入点：方向梯度直方图　出处：《哈尔滨理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：智能视频监控是计算机领域的一个新的应用方向和关注点,通过利用图像处理和机器视觉的相关方法,并分析监控视频,来实现对目标的定位、识别和跟踪,进一步实现对目标行为分析的目的,而目标跟踪技术作为智能视频监控的核心必将得到更广泛的应用。关于目标跟踪技术,关键问题是能够确保在当前帧中跟住目标,并在下一帧中不丢失目标。基于检测的目标跟踪算法相比单纯的目标跟踪算法表现出了更好的特性,因此本文将展开对目标检测和目标跟踪两个方面的研究,寻找更好的方法对目标进行描述,能够检测到目标并实时的跟住目标。对于目标检测系统,在其训练阶段,首先采用多尺度滑动窗口法对图像进行扫描和搜索,对每个滑动窗口分别计算方向梯度直方图(Histogram of Oriented Gradient,HOG)特征与语义局部二值模式(Semantic Local binary Pattern,SLBP)特征,然后利用主成分分析(Principal Component Analysis,PCA)将HOG特征降到特定的维数,并将两者结合构成联合特征,接着利用线性支持向量机(Support Vector Machine,SVM)训练分类器,最后采用上下文信息对检测到的目标框做优化处理。实验证明改进后的方法具有准确率高、受外界因素影响小、计算量小等优点。对于目标跟踪系统,本文采用的是经典的基于检测的跟踪-学习-检测(Tracking-Learning-Detection,TLD)跟踪框架。TLD算法在目标受到部分遮挡、变形、均匀光照等情况下能较好的完成跟踪任务,但是在光照变化不均、遮挡严重、跟踪目标模糊等情况下会出现误跟踪或跟踪失败等问题。针对存在的不足,本文将在检测模块中增加判别器,当跟踪目标具有较好的纹理特征时,采用SLBP分类器取代检测模块中的最近邻分类器,该分类器将图像转换为SLBP纹理特征向量,利用此向量对样本进行分类。实验表明,TLD-SLBP算法较其他的算法获得了更高的成功率,并通过在多个视频序列中测试体现出了较好的跟踪能力。
[Abstract]:Intelligent video surveillance is a new application direction and focus in computer field. By using the related methods of image processing and machine vision, and analyzing the monitoring video, we can locate, identify and track the target. Target tracking technology, as the core of intelligent video surveillance, will be used more widely. The key problem of target tracking technology is to ensure that the target can be followed in the current frame. In the next frame, there is no loss of target. The target tracking algorithm based on detection shows better characteristics than the pure target tracking algorithm. Therefore, this paper will carry out the research on target detection and target tracking. To find a better way to describe the target, it can detect the target and follow the target in real time. In the training stage of the target detection system, the multi-scale sliding window method is first used to scan and search the image. For each sliding window, the histogram of Oriented gradient histogram and the semantic local binary pattern are calculated, and then the principal component analysis (PCA) is used to reduce the HOG feature to a specific dimension, and to combine the two features to form a joint feature. Then the classifier is trained by linear support Vector machine (SVM), and the target frame detected is optimized by using context information. The experimental results show that the improved method has high accuracy and is less affected by external factors. For the target tracking system, the classical tracking-Learning-DetectionTLD-based tracking framework. TLD algorithm is partially occluded and deformed. Under the condition of uniform illumination and so on, the tracking task can be completed better, but in the case of uneven illumination, serious occlusion and vague tracking target, there will be some problems such as false tracking or tracking failure. In this paper, a discriminator is added to the detection module. When the target is tracked with good texture features, the nearest neighbor classifier in the detection module is replaced by the SLBP classifier, which converts the image into the SLBP texture feature vector. The experimental results show that the TLD-SLBP algorithm has a higher success rate than other algorithms and shows a better tracking ability by testing multiple video sequences.
【学位授予单位】：哈尔滨理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】