基于随机梯度提升决策树的行人检测算法设计与实现

发布时间：2018-06-28 04:17

本文选题：行人检测 + 区域建议网络　；参考：《浙江大学》2017年硕士论文

【摘要】：近年来,随着人工智能和机器学习的快速发展,计算机视觉也进入了发展的黄金时期,吸引了众多学者以及企业的目光。行人检测是计算机视觉中的重要课题之一,在智能视频监控和无人驾驶汽车等应用领域都有着举足轻重的地位。本文便着眼于行人检测这一重要且极具挑战的课题,行人检测本质上是个二分类问题,性能优异的行人检测算法既要有良好的分类算法也要有优秀的特征。本文的主要工作归纳如下:在行人检测领域中,已经被经常使用的分类算法有AdaBoost、支持向量机以及卷积神经网络中的Softmax分类函数等。梯度提升决策树(GBDT)是数据挖掘领域中性能非常出众的分类算法,在个性化推荐、金融预测等方面都有着成功的应用案例。然而,它目前还没有被应用于行人检测的算法中,因此本文的第一个创新点是把梯度提升决策树算法应用于行人检测中。本文设计了ACF/LDCF+GBDT算法,并在Inria、Caltech、Kitti几个主流的数据集上进行实验,实验结果证实了梯度提升决策树算法可以较好地适用于行人检测的研究中。卷积神经网络所得到的特征是对输入图像更抽象、更高层次的表达,高层次表达可以提升输入数据的区分度,我们采用一种优秀的卷积神经网络特征来进行行人检测算法的设计。FasterR-CNN中的区域建议网络(RPN)本身可以做为一个性能较好的行人检测器,但后面的分类器降低了其应有的性能。基于此本文提出了第二个创新点,先使用区域建议网络进行候选框的建议以及特征的提取,随后使用Bootstrapping策略分多个阶段采用梯度提升决策树算法进行模型的训练,充分挖掘疑似行人的负样本,并把这些样本加入训练集中的负样本里,从而逐步提升检测器的性能。此外,为了加快训练速度及有效地避免过拟合现象,我们采用了随机梯度提升的策略:每个阶段随机选取部分样本、随机选取部分特征用于决策树的训练,即训练过程中我们采用了随机梯度提升决策树算法。最终,本文设计了基于随机梯度提升决策树与区域建议网络的行人检测算法,并在当前流行的Caltech数据集上进行了实验。实验结果表明,经过以上改进后我们可得到一个性能非常优秀的行人检测器。
[Abstract]:In recent years, with the rapid development of artificial intelligence and machine learning, computer vision has entered a golden period of development, attracting the attention of many scholars and enterprises. Pedestrian detection is one of the most important subjects in computer vision. It plays an important role in intelligent video surveillance and driverless vehicle applications. This paper focuses on pedestrian detection, which is an important and challenging subject. Pedestrian detection is essentially a two-classification problem. The excellent pedestrian detection algorithm should have good classification algorithm as well as excellent characteristics. The main work of this paper is summarized as follows: in the field of pedestrian detection, the commonly used classification algorithms are Ada boost, support vector machine and Softmax classification function in convolution neural network. Gradient elevation decision Tree (GBDT) is a very outstanding classification algorithm in the field of data mining. It has been successfully applied in personalized recommendation and financial forecasting. However, it has not been applied to pedestrian detection, so the first innovation of this paper is to apply gradient lifting decision tree algorithm to pedestrian detection. In this paper, ACFR / LDCF GBDT algorithm is designed and tested on several main data sets of Inria Caltech Kitti. The experimental results show that the gradient lifting decision tree algorithm is suitable for pedestrian detection. The feature of convolution neural network is that the input image is more abstract and expressed at a higher level, which can improve the differentiation of input data. We use an excellent convolution neural network feature to design a pedestrian detection algorithm. The area recommendation Network (RPN) in FasterR-CNN can be used as a pedestrian detector with better performance, but the latter classifier reduces its performance. Based on this, a second innovation is proposed. Firstly, the proposed candidate and feature are extracted by using the regional suggestion network, and then the model is trained by gradient lifting decision tree algorithm in several stages using bootstrapping strategy. The negative samples of suspected pedestrians are fully mined and added to the negative samples in the training set to improve the performance of the detector step by step. In addition, in order to accelerate the training speed and avoid overfitting effectively, we adopt the strategy of random gradient lifting: random selection of parts of samples, random selection of some features for the training of decision trees in each stage. In the process of training, we adopt the stochastic gradient lifting decision tree algorithm. Finally, a pedestrian detection algorithm based on stochastic gradient lifting decision tree and regional recommendation network is designed and tested on the current popular Caltech data set. The experimental results show that after the above improvements, we can get a very good performance pedestrian detector.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】