【作者】 王焱；

【导师】 朱建科；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2017， 硕士

【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。更多还原

【Abstract】 With the rapid development of artificial intelligence and machine learning recently,the computer vision has entered into a period of prosperity.Pedestrian detection is one of the most important subjects in the computer vision,and it plays an important role in many applications,such as intelligent video surveillance and driverless car etc.In this paper,we concentrate on the challenging pedestrian detection subject,which is actually a binary classification problem.Both the classifier and features are crucial for a successful pedestrian detection algorithm and the principal studies of this paper are as follows:There are many classifiers having been used in pedestrian detection,such as Support Vector Machine,AdaBoost,and Softmax function in convolutional neural networks etc.Though Gradient Boosting Decision Tree(GBDT)is a widely used algorithm in data mining,personalized recommendation,financial forecasting for examples,it has not been used in the pedestrian detection.Thus the first novelty of this paper is to apply gradient boosting decision tree algorithm to the pedestrian detection,where the algorithm named ACF/LDCF+GBDT are designed and then tested on several datasets,including Inria,Caltech and Kitti.The experimental results show that gradient boosting decision tree can suit well for pedestrian detection.The features extracted from convolutional neural networks are abstract and high level descriptions of an input image.The high level descriptions are better at separating the input data,so we take advantage of these features to improve the pedestrian detection algorithm.The region proposal network(RPN)in Faster R-CNN is actually a good pedestrian detector and the classifier following it degrades its performance.Thus we propose the second novelty of this paper that we firstly use RPN to get the region proposals and features,and then adopt a multi-stage training by using GBDT algorithm with the bootstrapping strategy.Through mining hard negative examples and adding them into the training set gradually,the performance of pedestrian detector gets better and better.In addition,we use the stochastic gradient boosting to speed up the training process and avoid the overfitting.That is to say when training the decision tree model in each stage,a portion of samples and features are randomly selected.In a word,we use stochastic gradient boosting decision tree to train a pedestrian detector with the region proposals and features generated by RPN.Finally,we get a pedestrian detection algorithm combing stochastic gradient boosting decision tree and region proposal networks.We run several experiments on the leading pedestrian detection benchmark Caltech.The experimental results show that our pedestrian detector is very competitive.更多还原