Crowd density classification has been a challenging task in the field of computer vision, which has various applications in public and commercial domains. Many researches on the classification and recognition method of the crowd density have been introduced in the past, while there still exists the problems of inaccuracy, poor robustness and inefficiency. An adaptive crowd density classification method based on pixels and texture features is proposed in this paper. Core part of the method is to adopt different processing methods according to the corresponding crowd density. The method based on pixel regression method is used for the sparse crowd condition, while the texture features are applied in the dense crowd. Variety of texture features like local binary pattern (LBP), gray-level co-occurrence matrix (GLCM), Gabor, Haar-like and Wavelet group are used on the WorldExpo’10 dataset to obtain an optimum combination of these features, which is proposed to extract the texture features of the crowd images. Then the SVM classifier model based on Bayesian estimation is adopted to train the model which can filter the abnormal sample data to improve the accuracy and generalization performance of the algorithm. Meanwhile, a K-means clustering iterative training method based on optimized sorting samples is designed to improve the training speed in the training process. Extensive experiments from various aspects including parameter optimization, feature selection and model evaluation were conducted. The performance of the model is tested based on mean absolute error (MAE), mean squared error (MSE) and classification rate (CR) in dataset UCSD, Shanghai Tech_A and UCF_CC_50. The experimental results show that CR of the proposed method can reach to 98.2%, whose indexes of MAE and MSE also outperform the most existing methods. In general, the proposed approach in this paper has obvious advantages and great application value.

在计算机视觉领域，人群密度分类一直是一项具有挑战性的任务，计算机视觉在公共和商业领域具有各种应用。过去，人们对人群密度的分类和识别方法进行了许多研究，但仍然存在着不准确，鲁棒性差和效率低下的问题。提出了一种基于像素和纹理特征的自适应人群密度分类方法。该方法的核心部分是根据相应的人群密度采用不同的处理方法。在稀疏人群条件下使用基于像素回归的方法，而在密集人群中应用纹理特征。纹理的多种功能，如升OCAL二进制图案（LBP），灰色-级协同-发生矩阵（GLCM），伽柏，哈尔-样和小波组在WorldExpo'10数据集上使用来获得这些特征的最佳组合，建议将其提取出人群图像的纹理特征。然后采用基于贝叶斯估计的支持向量机分类器模型训练模型，该模型可以过滤异常样本数据，提高了算法的准确性和泛化性能。同时，设计了一种基于优化排序样本的K-means聚类迭代训练方法，以提高训练过程中的训练速度。从参数优化，特征选择和模型评估等各个方面进行了广泛的实验。基于数据集UCSD，Shanghai Tech_A和UCF_CC_50中的平均绝对误差（MAE），均方误差（MSE）和分类率（CR）对模型的性能进行了测试。实验结果表明，该方法的CR率可以达到98.2％，其MAE和MSE指标也优于大多数现有方法。总体而言，本文提出的方法具有明显的优势和很大的应用价值。