学术论文

      基于深度模型的场景自适应行人检测

      Scene adaptive pedestrian detection algorithm based on deep model

      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:
      To solve the problem that the detection effect of the existing machine learning based pedestrian detection algorithms decreases dramatically when the distributions of training samples and scene target samples do not match, a scene adaptive pedestrian detection algorithm based on the deep model is proposed.First, inspired by the Bagging(Bootstrap aggregating) mechanism, multiple relatively independent source samples are used to build multiple classifiers and then target training samples with confidence score are generated by voting.Secondly, using the automatic feature extraction ability of DCNN (deep convolutional neural network) and adding a deep auto-encoder to perform the source-target scene feature similarity calculation, a deep model-based scene adaptive classifier model is proposed and its training algorithm is designed.The experiments on the KITTI dataset demonstrate that the proposed algorithm performs better than the existing non-scene adaptive pedestrian detection algorithms.Besides, compared with the existing scene adaptive object detection algorithms, the proposed algorithm improves the detection rate on average by approximately 4%.
      作者: 蔡英凤 [1] 王海 [2] 孙晓强 [1] 袁朝春 [1] 陈龙 [1] 江浩斌 [2]
      Author: Cai Yingfeng [1] Wang Hai [2] Sun Xiaoqiang [1] Yuan Chaochun [1] Chen Long [1] Jiang Haobin [2]
      作者单位: 江苏大学汽车工程研究院,镇江,212013 江苏大学汽车与交通工程学院,镇江,212013
      年,卷(期): 2017, 47(4)
      分类号: TP391.4
      在线出版日期: 2017年8月15日
      基金项目: 国家自然科学基金资助项目,中国博士后基金资助项目,江苏省重点研发计划资助项目,江苏省自然科学基金资助项目,江苏省六大人才高峰资助项目