Considering the limitations of the feature extraction of high-density crowd images with perspective distortion, a crowd counting model, named LMCNN, that combines Global Feature Perception Network (GFPNet) and Local Association Feature Perception Network (LAFPNet) was proposed. GFPNet was the backbone network of LMCNN, its output feature map was serialized and used as the input of LAFPNet. And the characteristic that the Recurrent Neural Network (RNN) senses the local association features on the time-series dimension was used to map the single spatial static feature to the feature space with local sequence association features, thus effectively reducing the impact of perspective distortion on crowd density estimation. To verify the effectiveness of the proposed model, experiments were conducted on Shanghaitech Part A and UCF_CC_50 datasets. The results show that compared to Atrous Convolutions Spatial Pyramid Network (ACSPNet), the Mean Absolute Error (MAE) of LMCNN was decreased by 18.7% and 20.3% at least, respectively, and the Mean Square Error (MSE) was decreased by 22.3% and 22.6% at least, respectively. The focus of LMCNN is the association between the front and back features on the spatial dimension, and by fully integrating the spatial dimension features and the sequence features in a single image, the crowd counting error caused by perspective distortion is reduced, and the number of people in dense areas can be more accurately predicted, thereby improving the regression accuracy of crowd density.