基于人体典型部位特征组合的行人检测方法  

Pedestrian Detection Based on the Combined Features of Typical Parts of Human Body

作　　者：郭烈[1] 葛平淑[2] 赵一兵[1] 赵宗艳[1] 

机构地区：[1]大连理工大学汽车工程学院,大连116024 [2]大连民族学院机电信息工程学院,大连116600

出　　处：《汽车工程》2014年第4期496-502,共7页Automotive Engineering

基　　金：国家自然科学基金(61104165);高等学校博士学科点专项科研基金(20110041120024)资助

摘　　要：为实现汽车主动安全对行人的保护,提出一种组合人体典型部位特征的行人检测方法。利用查表型的Gentle Adaboost算法来训练经加权Fisher线性判别算法优化的梯度方向直方图特征,形成一个强分类器对行人腿部区域进行检测;行人初步定位后,根据其头部轮廓变动较小的特点,运用模板匹配的方法对其进行检测;最后根据腿部与头部检测结果,综合采用部位约束、特征转化和分类器阈值调整的方法对检测结果进行融合。结果表明,该方法能有效排除大部分虚警,提高检测行人的准确性。To realize the pedestrian protection of vehicle active safety, a pedestrian detection method based on combined features of human typical parts is proposed. Firstly, the look up table Gentle Adaboost algorithm is utilized to train the features of＇histogram of oriented gradient （HOG） having been optimized by weighted Fisher linear discrimination algorithm for creating a strong classifier to identify the region of pedestrian＇s legs. Considering the feature of its minor variation in contour, the head region is then detected with template matching method after the initial positioning of pedestrian. Finally the results of head and legs detections are fused by adopting a combination of parts constraints, feature transformation and classifier threshold adjustment. The results indicate that the method proposed can effectively exclude most of the false alarm, thus significantly enhance the detection accuracy of pedestrian.

关 键 词：汽车主动安全 行人检测 梯度方向直方图 模板匹配 特征组合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]