Crowd Counting via De-background Multicolumn Dynamic Convolutional Neural Network

The current state-of-the-art density map-based crowd counting methods have focused on designing convolution neural network (CNN)-based models to exploit multiscale features to handle crowd shape change due to perspective distortion. However, the significant concerns with such approaches are using static kernels and being not adaptive to input data. Again, the multiscale features should be more attentive towards background minimization. Hence, this chapter proposes a de-background multicolumn dynamic CNN for crowd counting to address the issues. The proposed model can handle crowd shape change due to perspective distortion and learn to minimize the background influence while doing crowd counting. Two benchmark crowd counting datasets, Mall and UCSD, are used to show the model’s effectiveness. In addition to this, a separate ablation study has been conducted to show the effect of individual modules of the proposed model.