A crowd density estimation method with multi-dimensional perspective multivariate information fusion（MDPMIF） is proposed for the problems that crowd density in two-dimensional images presents large differences with image viewpoint changes and multi-scale information loss in feature space. Firstly, the information of perspective change is encoded from ‘up-left-right-bottom’ direction, and the deep global contextual information is captured by progressive aggregation, and the scale relationship features of multi-dimensional perspective are extracted simultaneously. After that, a joint learning strategy is designed to obtain global scale relationship features and integrate global contextual expressions and global scale relationship features to obtain a more comprehensive description of perspective transformation. Finally, semantic embedding is used to realize the high and low order features to complement each other and enhance the quality of the output density map. Meanwhile, there are differences in crowd aggregation patterns in real scenes, and the simple density map method is prone to overestimate crowd counts for the low aggregation part of the image. Based on this, a high and low density multi-dimensional perspective multivariate information fusion crowd counting network（HLMMNet） is proposed on the basis of the MDPMIF network. A high and low density differentiation strategy is designed to adaptively divide the MDPMIF output into high and low density regions, keeping the MDPMIF network estimation results in the high density regions and using detection methods to achieve crowd counting correction in the low density regions, improving the robustness of the model. The experimental results show that the performance of this method is superior to other comparative methods.