CAJ | 학술논문

为了满足更快、更准、更鲁棒的行人检测需求，考虑交通监控视频图像质量不高与局部特征不明显的缺点，采用简单的行人特征来实现行人检测。除矩形度、高宽比、轮廓复杂度、宽度比、行人面积特征外，特定选用了对遮挡等干扰具有强鲁棒性的头部圆形度这一简单的局部特征。考虑交通监控视频图像中行人的尺寸变化，引入区域划分策略划分图像区域。兼顾高检测率和低误检率，根据分类误差最小原则与正样本分类率最大原则训练多个单特征多阈值 AdaBoost 行人检测器。为了优化多个行人检测器级联后的检测性能，在兼顾检测性能和检测速度的基础上，定义了以贡献率作为行人检测器的级联规则，依据贡献率大小确定的级联次序为基于高宽比、宽度比、矩形度、行人面积、轮廓复杂度和头部圆形度的行人检测器，依次进行级联，建立了新的多特征多阈值级联 AdaBoost 行人检测器。选用3个交通场景对行人检测器进行测试，并与单级 AdaBoost 行人检测器与现有2种级联 AdaBoost 行人检测器进行比较。分析结果表明：在3个交通场景的检测中，相比其他几种行人检测器，多特征多阈值级联 AdaBoost 行人检测器具有较高检测率、较快的检测速度和较低误检率，检测率最低为96．70％，误检率最高为0．67％，检测时间小于5 s，满足交通场景中对行人检测实时性和可靠性的要求。
위료만족경쾌、경준、경로봉적행인검측수구，고필교통감공시빈도상질량불고여국부특정불명현적결점，채용간단적행인특정래실현행인검측。제구형도、고관비、륜곽복잡도、관도비、행인면적특정외，특정선용료대차당등간우구유강로봉성적두부원형도저일간단적국부특정。고필교통감공시빈도상중행인적척촌변화，인입구역화분책략화분도상구역。겸고고검측솔화저오검솔，근거분류오차최소원칙여정양본분류솔최대원칙훈련다개단특정다역치 AdaBoost 행인검측기。위료우화다개행인검측기급련후적검측성능，재겸고검측성능화검측속도적기출상，정의료이공헌솔작위행인검측기적급련규칙，의거공헌솔대소학정적급련차서위기우고관비、관도비、구형도、행인면적、륜곽복잡도화두부원형도적행인검측기，의차진행급련，건립료신적다특정다역치급련 AdaBoost 행인검측기。선용3개교통장경대행인검측기진행측시，병여단급 AdaBoost 행인검측기여현유2충급련 AdaBoost 행인검측기진행비교。분석결과표명：재3개교통장경적검측중，상비기타궤충행인검측기，다특정다역치급련 AdaBoost 행인검측기구유교고검측솔、교쾌적검측속도화교저오검솔，검측솔최저위96．70％，오검솔최고위0．67％，검측시간소우5 s，만족교통장경중대행인검측실시성화가고성적요구。
In order to meet the practical demand for pedestrian detection with high speed,high accuracy and strong robustness,in view of the poor quality and unapparent local image features of traffic videos,some simple pedestrian features were chosen for pedestrian detection.Besides rectangle degree,ratio of height to width,shape complexity,normalized width,and pedestrian area,head density was applied because it is a simple local feature and has strong robustness for occlusion interference.Considering the size changing of pedestrian in the image,region division strategy was introduced into image region division.An improved training algorithm based on the minimum principle of classification error and the maximum principle of positive sample classification rate was implemented by considering both high detection rate and low false detection rate,thus several single-feature AdaBoost pedestrian detectors with multi-thresholds were obtained.To optimize the detection performance of cascade pedestrian detectors,the cascade rule was obtained in term of the contribution rate.The contribution rate was defined by weighing detection performance and detection speed.The cascade order was the detectors based on ratio of height to width,normalized width,rectangle degree,pedestrian area,shape complexity and head density.The pedestrian detectors were sequentially cascaded according to the cascade order,thus a cascade AdaBoost pedestrian detector with multi-features and multi-thresholds was constructed. The proposed pedestrian detector was tested by using 3 traffic scenes,and compared with single-cascade-level AdaBoost pedestrian detector and 2 existed cascade AdaBoost pedestrian detectors. Analysis result indicates that in 3 traffic scenes,compared with the other pedestrain detectors, the proposed pedestrain detector has higher detection rate,higher detection speed and lower false detection rate,the minimum detection rate is 96.70%,the maximum false detection rate is 0.67%,and the detection time is less than 5 s.So the detector satisfies the real-time and reliable requirements of pedestrian detection in traffic scene.1 tab,5 figs,24 refs.