ISSN 1000-1239 CN 11-1777/TP

• 论文 • 上一篇    下一篇


韩 光 赵春霞 胡雪蕾   

  1. (南京理工大学计算机科学与技术学院 南京 210094) (
  • 出版日期: 2009-11-15

An SVM Active Learning Algorithm and Its Application in Obstacle Detection

Han Guang, Zhao Chunxia, and Hu Xuelei   

  1. (College of Computer Science and Technology, Nanjing University of Science and Technology, Nanjing 210094)
  • Online: 2009-11-15

摘要: 障碍物检测是智能机器人要解决的非结构复杂环境感知的典型问题之一.在实际情况中,获得大量未标记样本是相对容易的,而标记这些样本则是极其繁琐和费时的工作,当前的研究工作很少涉及到这类问题的解决办法.将SVM主动学习算法引入到障碍物检测中,针对常规的SVM主动学习算法在应用中所遇到的问题和局限性,采用一种动态聚类过程来选取最有代表性样本和根据专家标记与当前SVM分类结果的差值来调整SVM超平面位置的两种策略对其进行了改进,提出了一种新的主动学习算法——KSVMactiv算法,并在真实的野外环境图像库上进行了实验.由实验结果可知:KSVMactiv算法仅用81个样本就能达到很高的检测效果,从而说明它能显著减少数据标记的工作量,且与已有主动学习算法相比收敛速度更快.

关键词: 智能机器人, 障碍物检测, KSVMactive, K均值聚类, 超平面位置校正

Abstract: Obstacle detection is one of the tasks which are solved for intelligent robot in the unstructured complicated environment perception. Large amounts of training data are usually necessary in order to achieve satisfactory generalization, and attaining these training data is also relatively easy. While manually labeling data is an expensive and tedious process. The current research work related to the solutions of the above problems is also very limited. Active learning algorithm is introduced to obstacle detection here. Aiming at the problems and limitations in the process of applying general active learning algorithm, two strategies are used to improve general SVM active learning algorithm. These two strategies use a dynamic clustering to select the best representative samples and, according to the difference of experts labeling and current SVM classification results, to tune the SVM hyperplane location. At the same time, a new SVM active learning algorithm is proposed, that is KSVMactive. Experiments are carried out in real wilderness environment image database. Experimental results demonstrate: very good detection results are obtained using KSVMactiv algorithm with only 81 samples, which can show that it can significantly reduce the workload of labeling data, and its convergence is better than other active learning algorithms.

Key words: intelligent robot, obstacle detection, KSVMactive, K-means clustering, hyperplane location correction