基于深度模型的场景自适应行人检测

发布时间：2018-03-24 11:23

本文选题：场景自适应　切入点：行人检测　出处：《东南大学学报(自然科学版)》2017年04期

【摘要】：针对现有基于机器学习的行人检测算法存在当训练样本和目标场景样本分布不匹配时检测效果显著下降的缺陷,提出一种基于深度模型的场景自适应行人检测算法.首先,受Bagging机制启发,以相对独立源数据集构建多个分类器,再通过投票实现带置信度度量的样本自动选取;其次,利用DCNN深度结构的特征自动抽取能力,加入一个自编码器对源-目标场景下特征相似度进行度量,提出了一种基于深度模型的场景自适应分类器模型并设计了训练方法.在KITTI数据库的测试结果表明,所提算法较现有非场景自适应行人检测算法具有较大的优越性;与已有的场景自适应学习算法相比较,该算法在检测率上平均提升约4%.
[Abstract]:In view of the shortcomings of the existing pedestrian detection algorithms based on machine learning, when the distribution of training samples and target scene samples mismatch, a scene adaptive pedestrian detection algorithm based on depth model is proposed. Inspired by Bagging mechanism, several classifiers are constructed from relative independent source data sets, and then automatic sample selection with confidence measure is realized by voting. Secondly, the feature extraction ability of DCNN depth structure is used. Adding a self-encoder to measure feature similarity in source-target scenarios, a scene adaptive classifier model based on depth model is proposed and a training method is designed. The test results in KITTI database show that, Compared with the existing scene adaptive learning algorithm, the proposed algorithm has more advantages than the existing non-scene adaptive pedestrian detection algorithm, and the average detection rate of the proposed algorithm is increased by about 4% compared with the existing scene adaptive learning algorithm.
【作者单位】：江苏大学汽车工程研究院;江苏大学汽车与交通工程学院;
【基金】：国家自然科学基金资助项目(U1564201,61403172,61601203) 中国博士后基金资助项目(2014M561592,2015T80511) 江苏省重点研发计划资助项目(BE2016149) 江苏省自然科学基金资助项目(BK20140555) 江苏省六大人才高峰资助项目(2014-DZXX-040,2015-JXQC-012)
【分类号】：TP18;TP391.41

【相似文献】