基于特征学习的目标检测与跟踪技术研究

发布时间：2018-01-14 07:47

本文关键词：基于特征学习的目标检测与跟踪技术研究　出处：《中国科学院大学(中国科学院光电技术研究所)》2017年硕士论文　论文类型：学位论文

【摘要】：目标检测与跟踪技术研究作为图像处理与模式识别领域的一个重要研究课题,有着非常广泛的应用前景,应用领域涉及公共安全、智能交通、视频检索、行为分析、人机交互等行业。本文在研究传统目标检测与跟踪算法的基础上,发现大多跟踪算法在跟踪过程中都因环境光照变化、目标姿态变化和被遮挡等导致跟踪失败。究其原因可能是由于选用的描述特征不具有良好的不变性或分类器选取不合理导致的。为此,本文围绕着如何选择合适的特征描述以及创建合理的分类器两个方面,就基于特征学习的目标检测与跟踪算法展开研究。基于特征学习的目标检测方面,本文详细介绍了基于Haar-like特征描述的Adaboost学习算法,这种算法通过构建二元分类器,利用分类器对目标与背景进行判定,其过程主要包括目标表示与训练学习分类器两个步骤。Haar-like特征之所以能够很好的描述目标是因为任意位置与尺度的图像块对不同种类Haar-like特征的响应(特征值)会不同,因此只要选择合适(尺度与位置)的特征组合便可完整、准确地表示目标;训练学习采用Adaboost算法将弱分类器组合成强分类器,提升分类器检测效果。该算法在目标检测中已经取得了较好的成果,但是由于特征不具有良好的不变性,致使算法的鲁棒性较差。文中首先采用LBP算子对图像进行重新编码,然后使用归一化Haar-like特征表示目标,新的特征描述具有更好的光照和尺度不变性,使目标检测算法能够适应更多的应用场景。实验结果表明经过特征改进后的算法鲁棒性更强、检测效果也得到进一步提升。基于特征学习的目标跟踪方面,由于基于特征学习的目标检测算法是通过离线学习的方法训练检测子,此方法训练耗时且仅针对特定目标进行检测,很显然不能基于这样的检测流程对目标进行跟踪。考虑到多实例在线学习(MIL)能够较好地解决基于检测的跟踪算法由于长时间处于无监督学习状态导致模板漂移的问题,文中将其引入跟踪框架中。该算法通过在线采样的方式构造训练样本,采用Haar-like特征对目标进行描述,基于Online boosting学习算法选择分类器,通过实验发现其能够适应大多数简单场景,但是在复杂环境下算法表现欠佳。究其原因可能出现在目标表示和分类器学习上。首先特征描述缺乏不变性;其次是为满足实时性要求,算法依赖的弱特征(弱分类器)池中弱分类器数量本身就有限,而且原算法采用随机选取的方法,选择弱分类器时存在一定的盲目性;再者通过线性组合弱分类器得到强分类器,相当于每个弱分类器对最终检测子的贡献一样,与实际情况不符。因此文中采用归一化的Haar-like特征和类Adaboost的算法选择分类器来实现目标跟踪。为验证本文算法的有效性,文中采用大量数据集进行实验论证,同时与跟踪效果与时效性都比较不错的MIL、CT算法进行了对比实验。该方法不仅能够适应更多复杂场景,同时也能在一定程度上适应目标微小形变、旋转和被遮挡的情况。但该算法也存在着不足,当目标被严重遮挡或消失后重现算法失效。
[Abstract]:An important research topic in research on target detection and tracking as image processing and pattern recognition, has a very broad application prospects and application fields of public safety, intelligent transportation, video retrieval, behavior analysis, human-computer interaction and other industries. Based on the detection and tracking algorithm of traditional targets, found mostly in tracking the algorithm in the tracking process because of the change of illumination, pose variations and occlusion cause tracking failure. The reason may be due to the selection of features does not have good invariance or classifier selection unreasonable. Therefore, this paper focuses on how to choose the appropriate description of the features of the two aspects and create a reasonable the classifier, feature learning algorithm based on target detection and tracking is researched. Characteristics of learning on target detection, this paper introduces in detail based on Ha Adaboost learning algorithm AR-like description, this algorithm constructs two element classifier of target and background are determined by the classifier, which mainly includes object representation learning classifier two steps.Haar-like features and training can be a good description of the target image block is because any position and scale of the response characteristics of the different kinds of Haar-like (eigenvalues) will be different, so long as appropriate (position and scale) combination of features can be complete, accurate representation of the target; trained by Adaboost algorithm to weak classifiers combined into a strong classifier, improve the detection effect of classifier. This algorithm has achieved good results in target detection, but not due to the characteristics of has a good invariance, resulting in the poor robustness of the algorithm. Firstly, LBP operator is used to re encoding of the image, then use. Haar-like said the target feature, the new feature description has better illumination and scale invariance, the target detection algorithm can adapt to more scenarios. The experimental results show that the algorithm has better robustness characteristics improved, detection effect has also been further improved. Based on the characteristics of the learning object tracking, due to the characteristics of the learning goals detection algorithm based on the method is through the off-line learning training test, this training method is time-consuming and only for specific target detection, obviously not based on this detection process for target tracking. Considering the multiple instances of online learning (MIL) can be used to solve the tracking algorithm based on the detection of the long time in unsupervised learning state result of template drift problem, this paper will introduce the tracking framework. Construct training samples through the algorithm of online sampling, using Haa To describe the target R-like feature, Online boosting learning algorithm based on classifier selection, found through experiments which can adapt to the most simple scene, but in the algorithm under complex environment is poor. The reason for this may appear in the target representation and classifier learning. First description of the features of lack of invariance; second is to meet the real-time requirements of weak feature algorithm dependent (weak classifiers) the number of weak classifier pool itself is limited, but the original algorithm uses methods, the blindness of selecting certain weak classifiers; and a strong classifier by a linear combination of weak classifiers, each weak classifier is equivalent to the final detection sub contribution, inconsistent with the actual situation. So the Haar-like features and Adaboost normalization algorithm selection classifier to realize target tracking. In order to verify the method of The effect of a large number of data sets is demonstrated in the paper and the tracking effect and timeliness are relatively good MIL, CT algorithm are compared. The method can not only adapt to more complex scenes, but also can adapt to the small deformation target in a certain extent, rotation and obscured. But the algorithm there are also some shortcomings, when severe occlusion or disappeared after reconstruction failure.

【学位授予单位】：中国科学院大学(中国科学院光电技术研究所)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】