Crowd Visualization and Counting by Smooth Dilated Convolutional Network

Abstract

With mass urbanization, culturally diversified country with many cultural and religious gatherings happening time and again, the need of crowd estimation from single image and information on the distribution of crowd in the same is deemed necessary. The old-fashioned way of keeping records, sensor-based counting fails when the crowd movement is dynamic and/or random. Task is challenging due to geometric distortion, perspective distortion, severe occlusion, illumination condition in the image. The forementioned challenges has been addressed by Deep Learning Convolutional Neural Network where in CNN is employed as a feature extractor and Smoothed Dilated CNN is used in the backend, that facilitates aggregation of multi-scale contextual information by increasing the receptive field with same resolution removing the gridding artifacts. Model is end-to-end trainable since it employs pure convolutional structure and can accept arbitrary size and resolution of input image for conversion into density map which is used for crowd counting. Training of the model begins with the generation of ground-truth density map which is computed based on geometry-adaptive kernel to account for perspective effect on the denser crowd and fixed kernel on the sparse crowd. ShanghaiTech dataset is been used which comprises of 1198 tagged images with a total amount of 330,165 persons. Comparison between dilation rate 2 and 4 for both Part_A and Part_B of ShanghaiTech dataset is made. Upon evaluation of the model with the Csrnet where smoothing of the dilated convolution is not implemented, the counting accuracy and quality of the density map for both Part_A and Part_B of the dataset has been significantly increased.