魏文戈 陈舒荻

摘   要：在计算机视觉领域，针对小角度俯拍下的站台人群计数的研究工作较少，且计数精度普遍较低。人群计数算法往往通过图像分割识别出图片中的所有行人个体，并进行数量统计，具有很重要的现实意义。然而现有的图像分割算法往往只能适用于简单场景下的简单分割任务。由于小角度俯拍下的站台场景中存在行人近大远小、行人互相遮挡和行人轮廓姿态多样等原因，因此给有效分割计数带来了较大的挑战。针对这一任务，提出了距离自适应卷积神经网络（Distance Adaptive Convolutional Neutral Network，简称DACNN），通过改进回归对象和设计距离自适应卷积层，成功实现了对小角度俯拍下站台人群的准确计数。经过一系列实验表明，该模型不仅计数精度高，而且计数速度较快、鲁棒性良好，具有广阔的运用前景。

关键词：小角度俯拍;站台人群计数;图像分割;距离自适应

中图分类号：TP391                                      文献标识码：A

Platform Crowd Counting Under Small Angle

WEI Wen-ge？覮，CHEN Shu-di

（College of Computer Science and Technology，Nanjing University of Aeronautics and Astronautics，Nanjing，Jiangsu 211106，China）

Abstract：In the field of computer vision，there are few research work on platform crowd counting under small angles and the counting accuracy is generally low. Crowd counting algorithms often identify all the individuals in the image through image segmentation，and make the counting of the crowd，which is of great practical significance. However，the existing image segmentation algorithms are often only applicable to simple segmentation tasks in simple scenes. Since the human body caused by small angles in the vicinity looks larger while in the distance seems small，the body block each other and have diverse stance，to effectively split the statistics becomes a great challenge. In response to this task，this paper proposes Distance Adaptive Convolutional Neutral Network （DACNN）. By designing a distance-adaptive convolutional layer and improving the model loss function，we can achieve a better performance under small angles of platform crowd counting. After a series of experiments show that the algorithm model not only has a high counting accuracy and fast counting speed，but also has good practicality，can be widely used.

Key words：small angle;platform crowd counting;image segmentation;distance-adaptive

人群计数，作为图像识别的一个重要研究领域，通过图像分割识别出图片中的所有人群个体，并进行数量统计，可广泛运用于安防等领域。目前针对小角度俯拍下的站台人群计数的研究工作较少且准确率较低，不能较好地应用于复杂的实际场景中。由于监控摄像机小角度俯拍导致的站台行人近大远小，行人互相遮挡和輪廓姿态多样等问题，给有效分割计数带来了较大挑战。针对这一问题，提出了距离自适应卷积神经网络（Distance Adaptive Convolutional Neutral Network），对传统回归对象进行了改进，并创新性地设计了距离自适应卷积层，从而实现了对小角度俯拍下的站台人群计数的有效计数。……