Crowd Counting and Localization in Haze and Rain

Adverse weather conditions such as haze and fog often significantly reduce the performance of crowd counting models. An intuitive solution is to preprocess degraded images by applying image restoration techniques prior to crowd counting. However, this solution introduces additional computational complexity and may produce restored images with noise and artifacts that is harmful to the subsequent crowd counting task. To mitigate the two issues, we integrate an image restoration module (IRM) into a unified framework to propose an effective network for crowd counting and localization in haze and rain. The lightweight IRM is designed to guide the network to learn haze-aware knowledge in feature space, which is removed in the inference phase without increasing the computational cost. In addition, two new datasets are constructed to evaluate the crowd counting methods in haze and rain. Extensive experiments on both synthetic and real-world datasets demonstrate the superiority of the proposed method. The code is available at https://github.com/lizhangray/Dehaze-P2PNet.

I. Introduction

Crowd counting is a fundamental task in computer vision that aims to estimate the number of people in images. Most existing methods [1]–[5] generate intermediate representations of learning targets, such as density maps, where the crowd count is computed via summing over the estimated density map. However, numbers counting alone can hardly support the downstream tasks based on the crowd distributions. In response to this problem, numerous methods have been proposed for more challenging fine-grained prediction of the exact locations of individuals. Specifically, some approaches [6], [7] bypass the error-prone steps and directly predict the center points of heads, yielding encouraging counting performance and impressive localization accuracy.