基于深度优先随机森林分类器的目标检测  

Object detection for depth-first random forest classifier

作　　者：马娟娟 潘泉[1] 梁彦[1] 胡劲文[1] 赵春晖[1] 王华夏[1] MA Juanjuan, PAN Quan, LIANG Yan, HU Jinwen, ZHAO Chunhui, WANG Huaxia(Key Laboratory of Information Fusion Technology Ministry of Education, School of Automation, Northwestern Polytechnical University, Xi＇an 710129, China)

机构地区：[1]西北工业大学自动化学院信息融合技术教育部重点实验室,西安710129

出　　处：《中国惯性技术学报》2018年第4期518-523,共6页Journal of Chinese Inertial Technology

基　　金：国家自然科学基金重点项目(61135001);国家自然科学基金项目(61473230;61603303);陕西省自然科学基础研究计划项目(2017JQ6005;2017JM6027);爱生创新发展基金项目(ASN-IF2015-1502);中央高校基本科研业务费资助项目(3102017jg02011)

摘　　要：从机载视觉传感器获取的图像中检测近距离目标,对小型无人机飞行安全非常重要,需要大量样本训练分类器以提高目标检测的准确性。然而,如果训练样本太大,随着树的层数增加,广度优先方法训练随机森林分类器会导致欠拟合问题。针对这个问题,提出了深度优先方法递归训练随机森林分类器,每次递归过程只分裂一个节点。实验表明,在SenseAndAvoid数据集目标检测的平均准确率是69.3%,比广度优先方法训练的随机森林分类器高7.6%。深度优先方法递归训练随机森林分类器,能有效抑制广度优先方法训练时的欠拟合问题,提高了随机森林分类器的泛化能力和目标检测的准确性。The ability to detect the visible objects from the images obtained by onboard vision sensors is very important for flight security of small unmanned aerial vehicle. A great number of samples are needed to train the classifier to improve the precision of object detection. However, breadth-first random forest classifier training will lead to underfitting when the number of tree layers increases. To solve this problem, depth-first is injected into random forest classifier for implementing the tree training, where only one node is split at each recursive time. Experiments demonstrate that the detection average precision on SenseAndAvoid dataset is 69.3%, which improves the average precision by more than 7.6% compared with that of breadth-first random forest classifier training. Depth-first random forest classifier training is able to effectively inhibit underfitting, which improves the generalization performance of random forest classifier and the precision of object detection.

关 键 词：无人机 目标检测 深度优先 随机森林分类器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]