Abstract: Crowd counting is difficult to get accurate statistics due to shading, shadows and changes in crowd density. This paper presents an approach to combine the convolutional neural networks and density features map legitimately. We segment the crowd scene into many blobs according to the density. For low-density blobs, Retinex algorithm is used to denoise the scene and then the scene is transformed into HSV color space to locate the pedestrian. Convolutional neural networks are used to extract the pedestrian features with grid loss function to avoid the occlusion issue. For high density blobs, crowd density distribution features are extracted to train the multiple kernel regression models to estimate the numbers. Experiments are conducted on datasets PETS2009, UCSD. The experimental results show that the proposed method improves the accuracy to some extent in comparison with other algorithms.