基于Mask R-CNN的行人分割

Pedestrian Segmentation Based on Mask R-CNN

在安防、无人驾驶、军事等领域,辨别行人是一项基本任务,在无法获得行人面部特征的情况下,步态等生物信息也可作为识别行人的依据。采用行人实例分割的方法,获取高精度的行人分割掩码,有利于寻找行人腿部轮廓与步态之间的关系。基于Mask R-CNN模型,针对行人的长宽比例,调整RPN网络,剔除宽高比大于1∶1的部分,并使用扩充的行人实例分割Penn-fudan数据集进行迁移学习,行人分割掩码的交并比(IoU)较预训练模型提高了9%,获取了更高精度的行人分割掩码以及腿部轮廓。

In the fields of security,driverless and military,it is a basic task to identify pedestrians.In the absence of facial features of pedestrians,biological information such as gait can also be used as a basis for identifying pedestrians.In this paper,the pedestrian segmentation method is adopted to obtain high-precision pedestrian segmentation mask,which is helpful to find the relationship between pedestrian leg contour and gait.Based on the Mask R-CNN model,the average length-to-width ratio of pedestrians in the needle is adjusted,the RPN network is adjusted,the part with aspect ratio greater than 1∶1 is removed,and the Penn-fudan data set is segmented using the extended pedestrian instance for migration learning.Pedestrian segmentation The mask's IoU value is 9%higher than the pre-training model,and a more accurate pedestrian segmentation mask and leg contours are obtained for preparation for subsequent gait recognition.