【摘要】 密集场景下准确人群计数和定位，对于保障公共安全具有重要的意义。针对密集人群计数与定位易受人群分布不均、背景干扰等因素的影响，导致计数定位不准确的问题，提出一种基于区域感知校准的自适应人群计数与定位方法。通过构建金字塔结构提取人群图像的多尺度特征，增强特征关联性，并设计可变形几何自适应模块学习不同分布的人群几何特征，以增强对人群分布不均的适应性。在此基础上，提出区域感知和区域校准模块，提取全局上下文特征和区域特征，克服了背景干扰造成的定位与计数不准问题。接着通过双分支卷积预测通路，输出生成点的预测位置和置信度分数，以提高网络的定位与计数精度。最后提出改进二分图最大匹配Hopcroft-Karp算法对真值点与预测点进行匹配校准，从而完成人群定位与计数。实验结果表明，所提方法分别在公开的ShanghaiTech Part A和Part B数据集、NWPU-Crowd数据集、UCF-QNRF数据集上评价指标均优于对比算法，且定位精度较P2Pnet分别提高了3.5%、6.1%、11.3%和8.1%,能够有效提高人群定位与计数的准确度。更多还原

【Abstract】 Realizing accurate crowd counting and positioning in dense scenarios is of great significance for ensuring public safety. Addressing the issue of inaccurate counting and localization due to uneven crowd distribution and background interference, this paper proposed an adaptive crowd counting and location method based on area perception calibration. By building a pyramid structure to extract multi-scale features of crowd images and enhance feature correlation, a deformable adaptive geometry module was designed to learn geometric features of different crowd distributions, in order to enhance adaptability to uneven crowd distribution. Based on this, a regional awareness and calibration module was put forward to extract the global context and regional characteristics, overcoming the inaccuracies in localization and counting caused by background interference. Subsequently, double branch convolution prediction pathway was used to output predicted positions and confidence scores of generated points, in order to improve the precision of the network in localization and counting. Finally, an improved binary chart Hopcroft-Karp maximum matching algorithm was put forward for true value calibration points and prediction points matching, thus completing crowd localization and counting. The experimental results show that the proposed method outperforms contrast algorithms respectively in open ShanghaiTech Part A and Part B dataset, NWPU Crowd dataset, UCF QNRF dataset in terms of evaluation indexes, with an improvement in localization precision of 3.5%, 6.1%, 11.3% and 8.1% respectively over that of P2Pnet, which can effectively improve the accuracy of crowd localization and counting.更多还原