基于G-UNet的多场景行人精确分割与检测

Pedestrian segmentation and detection in multi-scene based on G-UNet

目前的语义分割方法可以得到行人的轮廓,但在行人相互遮挡时,无法直接得到图中行人的数量、身高和中心位置等信息。针对这一缺陷,提出了G-UNet模型算法:在语义分割主干之外,增加一个行人区域的高斯椭圆密度核检测分支,通过核的极大值点、垂直和水平轴尺度,分别检测行人的中心位置、高度和宽度,并由密度核极大值点的唯一性,解决了行人遮挡的检测难题。另外,UNet以空间对称的方式将底层和高层特征进行硬性的拼接,使得50%的固定误差直接传播到底层。本文提出可训练的柔性系数拼接方式,可以得到最优的误差分配传播方式。最后,传统的损失函数的误差值与行人标定面积成正比,导致小尺度行人容易漏检,本文提出目标增强损失函数提高网络检测小尺度行人的能力。在自建行人分割数据库中,实验结果证明了本文方法的有效性且优于其他方法。

Current semantic segmentation methods can obtain the outline of pedestrians, but when pedestrians block each other,the information such as the number,height and central position of the pedestrians in the figure can not be directly obtained. To solve this problem,we propose the G-UNet model algorithm that,in addition to the semantic segmentation backbone,a Gaussian ellipse density kernel detection branch of the pedestrian area is added,so that the center position,height and width of the pedestrian are detected respectively by the maximum point of the kernel,vertical axis and horizontal axis scale. The uniqueness of the maximum point of density kernel solves the detection problem of pedestrian occlusion. Besides,UNet stiffly concatenates the features of the bottom layer and the top layer in a spatial symmetric way so that 50% fixed errors are directly transmitted to the bottom layer. Then,we propose a trainable Soft Connection to obtain the optimal error distribution propagation method. Finally,to solve the problem that the value of the traditional loss function is proportional to the demarcated area of pedestrian,which makes small-scale pedestrians easy to be undetected,an Objective Enhanced Loss was proposed to improve the ability of detecting small-scale pedestrians of network. In the self-built pedestrian segmentation dataset,the experimental results show that the proposed method is effective and superior to other methods.