概率图车载热成像行人检测RoIs提取方法

道路场景复杂、热成像纹理信息较少以及图像品质不稳定，RoIs提取面临挑战.阈值分割RoIs提取更多关注行人局部细节和像素间邻域关系，容易产生行人遗漏、背景粘连和行人断裂，且很难控制RoIs总量.模拟人类视觉，关注图像显著性区域及其位置和大小，提出概率图RoIs提取方法，设计凸-凹形曲线映射像素灰度值增强图像对比度；基于图像签名方法获取显著性图.融合灰度强度和显著性概率图并从中提取图像前景；设计算法搜索路面估计限定的概率图区域生成RoIs.实验表明，相对阈值分割，本文方法能够提高RoIs定位准确度、控制RoIs总量和显著减少非行人RoIs；提取等量单帧RoIs，召回率提高不低于9%.

RoIs Extraction Method Using Probability Map for Vehicle-Mounted Thermal Imaging Pedestrian Detection

RoIs extraction is a challenging task due to scene complexity,less image texture and unstable image quality.Paying more attention to local pedestrian details and neighborhood relations of pixels,threshold segmentation is prone to give rise to mis-segmentation such as adhesion,omission,breakage and uncontrollable RoIs amount.Imitating human’s eyes,we focus on an image saliency region,size and location and propose a new RoIs extraction method using probability map.Design convex-concave curves to map an image pixel gray for enhancing the image contrast;get a saliency map based on image signature manner;fuse intensity and saliency probability images and then extract the image foreground;design an algorithm to search the probability map region limited by a road horizon and generate RoIs.Experimental results show that our method can improve RoIs locating accuracy,control RoIs amount and reduce non-pedestrian RoIs when comparing with threshold segmentation method.Our recall increases no less than 9% when same RoIs amount are extracted.