自然场景下基于部分模型的车体检测  

Vehicles detection in natural scene based on part model

作　　者：陈帅[1] 张勇[1] 牛常勇[1] 

机构地区：[1]郑州大学信息工程学院,河南郑州450004

出　　处：《计算机工程与设计》2015年第4期1090-1093,共4页Computer Engineering and Design

基　　金：国家自然科学基金项目(60901078);郑州市科技领军人才基金项目(10LJRC189)

摘　　要：为检测车体位置和分类车辆的类型,运用一种可重构模型来实现车体的检测。使用由一个表示整个模型的低分辨率模板和多个表示各个部分的高分辨率模板组成的模型完成自然场景下车体的检测,为得到该模型的最优配置,提出一种部分模型划分算法确定构成模型的模板的位置和大小,有效提高检测的召回率,通过支持向量机算法得到各个模板的最优参数配置。在检测算法中提出目标融合算法,避免多次检测到同一目标。实验结果表明,对比PBT(part basis template model)的方法,该方法改善了检测召回率和分类准确率。Aiming to design a reconfigurable model to deal with vehicle detection and classificationm,the model that consisted of one low-resolution template covering the whole model and several small and high-resolution templates representing the components was proposed to detect vehicles in natural scene.To achieve the optimal configuration of this model,a methodology of divided part model algorithm was put forward to define the positions and sizes of the templates,and this methodology significantly improved the accuracy of vehicles detection.The optimal configuration of templates was obtained using the SVM algorithm.Besides,an object fusion algorithm was used in detection to avoid the phenomenon that one target was detected repeatedly.The experiment result shows that compared with the PBT（part basis template model）method,the proposed method promotes both the recall rate and the accuracy in vehicle detection and classification.

关 键 词：人工智能 车体检测 可重构模型 支持向量机 部分模型 目标融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]