智能车多类障碍物贝叶斯网络分类方法

Multi-class Obstacle Classification Using Bayesian Network for Intelligent Vehicles

本文提出了一种用于智能车,基于贝叶斯网络分类模型的道路多类型障碍物分类方法.以立体视觉方法获取的目标的三维尺寸为特征,采用预设经验的NPC算法训练产生优化的网络结构,通过EM算法训练产生离散节点的条件概率表和连续节点的条件概率分布,产生贝叶斯网络分类模型.将检测到的道路目标分成行人、骑行者、小汽车、小货车、卡车五类.用公共图像数据库KITTI对本方法进行测试,实验结果表明本文所提方法优于现有同类工作.

This paper presented a method of classifying multi-class obstacles for intelligent vehicles. Bayesian Network was used to establish a classification model,and the three-dimensional sizes of the target detected by the stereo vision method were used as the features. The model was trained by using Necessary Path Condition( NPC) algorithm restrained by presupposition experience to determine the optimized structure and by using Expectation Maximization( EM) algorithm to obtain the conditional probability table of the discrete nodes and the conditional probability distribution of the continuous nodes. The targets were divided into five classes including pedestrian,riders,cars,vans and trucks. Experiments were conducted on the KITTI public database,and the results show that the proposed method outperforms the existing similar works.