In order to solve the problem of uneven distribution of crowd and massive network learning parameters, a method for accurate high-density crowd counting was proposed, which is composed of Pixel-level Attention Mechanism (PAM) and improved single-column crowd density estimation network. First of all, the PAM was used to generate a high-quality local crowd density map by classifying the crowd images at pixel level, and the Full Convolutional Network (FCN) was used to generate the density mask of each image, and the pixels in image were divided into different density levels. Then, using the generated density mask as the label, the single-column crowd density estimation network was used to learn more representative features with fewer parameters. Before this method was proposed, the counting error of Network for Congested Scene Recognition (CSRNet) method was the smallest on part_B of Shanghaitech dataset, the UCF_CC_50 dataset and the WorldExpo'10 dataset. Comparing the error results of proposed method with CSRNet, it is found that this method has the Mean Absolute Error (MAE) and Mean Squared Error (MSE) on part_B of Shanghaitech dataset reduced by 8.49% and 4.37%; the MAE and MSE on UCF_CC_50 dataset decreased by 58.38% and 51.98% respectively, which are of significant optimization, and the MAE of overall average value on the WorldExpo'10 dataset reduced by 1.16%. The experimental results show that when counting the unevenly distributed high-density crowd, the method of combination of PAM and single-column crowd density estimation network can effectively improve the accuracy and training efficiency of high-density crowd counting.