摘要
针对交通道路中的目标检测问题,提出了一种基于锚点聚类、全锚点训练策略及强化交并比(SIoU)的交通目标检测方法(T-Faster RCNN)。首先,通过一个基于交并比距离的K-means聚类获取交通目标的宽高在比例和尺度两个几何属性的先验知识,生成锚点边界框;其次,将分类损失与焦点损失相结合进行全锚点训练;再次,基于两个边界框所构成的最小闭包生成SIoU,用于筛选建议。在KITTI数据集上进行的对比实验表明本文方法比Faster RCNN的mAP提高了14.4%。
In this paper,a traffic object detection method(T-Faster RCNN),which involves its improvement in three strategies,is proposed.Firstly,an IoU-based K-means clustering strategy is used to generate anchor boxes,where the quantitative geometric attributes of the aspect ratios and scales are extracted from all the objects in the training dataset.After retaining all the anchors,the RPN classification loss borrowed from focal loss is second employed to train the anchors.Finally,a strengthened IoU,which is induced from the minimum closure of two bounding boxes,is proposed to filter proposal candidates.The comparison experiments were implemented on KITTI dataset.In all,the mAP of T-Faster RCNN,which applies the three strategies,is 14.4%higher than that of Faster RCNN.
作者
张琦
丁新涛
王万军
周文
ZHANG Qi;DING Xingtao;WANG Wanjun;ZHOU Wen(School of Computer and Information,Anhui Normal University,Wuhu 241003,China;Anhui Provincial Key Laboratory of Network and Information Security,Wuhu 241003,China)
出处
《皖西学院学报》
2019年第5期50-55,共6页
Journal of West Anhui University
基金
安徽省自然科学基金面上项目(1808085MF171)