扩展目标的稳定检测技术研究

发布时间：2018-05-08 18:32

本文选题：目标检测 + 扩展目标　；参考：《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

【摘要】：目标检测作为图像处理、识别、跟踪的一个重要组成部分,一直都是人们研究的热点。在视频监控、车辆导航、目标跟踪等多领域内都有着广泛的应用前景。传统的目标检测方法主要是基于特征点的检测,在复杂场景中存在漏检率高、检测率低等问题。本文研究的主要是扩展目标的检测,此类目标通常在视场中会占据较大位置,同时会有较大的姿态、尺度变化以及遮挡等问题。而利用传统的基于特征点匹配的检测算法,很有可能导致特征匹配失败,影响检测率,从而也会影响后续应用,如目标跟踪。因此研究一种较为鲁棒、稳定的目标检测算法具有重要意义。针对扩展目标检测,本文详细分析常用的局部特征提取方法包括SIFT、HOG、Harr-like、LBP等,对HOG特征进行了深入分析并改进,针对HOG特征不能适应目标尺度变化问题,构建了特征金字塔,使HOG特征具有尺度不变性。针对提取的HOG特征维数过多,计算量大的问题,采用了主成分分析(PCA)进行特征降维。利用支持向量机(SVM)对样本图像进行训练,实现了HOG+SVM目标检测算法,并针对该算法在目标出现遮挡及较大姿态变化时,容易出现漏检及误检情况,从部件的角度构建了基于部件的目标检测算法。在对部件模型参数的训练过程中,由于输入样本只对目标整体进行了标注,并没有对各个部件进行标注,因此利用了Latent-SVM进行训练。为了进一步提高算法的鲁棒性,在模型参数训练过程中,首先对样本进行了聚类,根据聚类的结果构建了多视角混合训练模型,实现了混合多部件的目标检测算法。实验表明,该算法在物体出现形变、遮挡及姿态变化时,能够准确检测到目标。尽管利用部件的思想,对目标进行检测,能够提升检测精度,但是增加了模型的复杂度,检测速度降低。针对此问题,本文结合了级联检测的思想,通过利用级联检测,能够快速的过滤掉图像中的非目标部分,实验表明,利用级联思想在保证检测精度不下降的情况下,能够加快算法的运行效率,缩短检测时间。
[Abstract]:As an important part of image processing, recognition and tracking, target detection has always been a hot topic. Video surveillance, vehicle navigation, target tracking and other fields have a wide range of applications. The traditional methods of target detection are mainly based on feature points. In complex scenes, the detection rate is high and the detection rate is low. This paper focuses on the detection of extended targets, which usually occupy a large position in the field of view, at the same time, there are many problems, such as attitude, scale change and occlusion. However, the traditional detection algorithm based on feature point matching may lead to the failure of feature matching, which will affect the detection rate, and will also affect the subsequent applications, such as target tracking. Therefore, it is important to study a robust and stable target detection algorithm. Aiming at the extended target detection, this paper analyzes in detail the commonly used local feature extraction methods, such as sift hog Harr-like LBP, and analyzes and improves the HOG feature in depth. Aiming at the problem that the HOG feature can not adapt to the change of target scale, the feature pyramid is constructed. The HOG features are scale-invariant. In order to solve the problem that the extracted HOG feature dimension is too many and the computation is large, the principal component analysis (PCA) is used to reduce the feature dimension. The HOG SVM target detection algorithm is realized by using support vector machine (SVM) to train the sample image. The algorithm is prone to miss detection and false detection when the target appears occlusion and larger attitude change. An object detection algorithm based on components is constructed from the point of view of components. In the training process of component model parameters, because the input sample only annotates the whole target, and does not annotate each component, the Latent-SVM is used to train. In order to further improve the robustness of the algorithm, in the process of model parameter training, the samples are first clustered, and the multi-view hybrid training model is constructed according to the result of the clustering, and the target detection algorithm of mixed multi-component is realized. Experiments show that the algorithm can detect the target accurately when the object appears deformation, occlusion and attitude change. Although the detection accuracy can be improved by using the idea of components, the complexity of the model is increased and the detection speed is reduced. Aiming at this problem, this paper combines the idea of cascade detection, by using cascade detection, can quickly filter out the non-target part of the image. The experiment shows that the concatenation method can ensure the detection accuracy does not decrease. It can speed up the efficiency of the algorithm and shorten the detection time.
【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】