Crowd density estimation based on face detection under significant occlusions and head pose variations

Abstract

Counting and detecting occluded faces in the crowd is a challenging task in computer vision. In this paper, we propose a new approach to crowd estimation based on face detection under significant occlusion and head pose variations. Most of the state-of-the-art face detectors are unable to detect occluded faces. To address the problem, a novel approach for training various detectors is described. In order to obtain a reasonable evaluation of our solution, we trained and tested the model on our substantially occluded data-set. Images of the face up to 90$${^\circ }$$ out-of-plane rotation and the faces with 25%, 50%, and 75% occlusion level. In this study, we trained the proposed model on 48,000 images obtained from our data-set consisting of 19 crowd scenes. To evaluate the model, 109 images with the face counts ranging from 21 to 905 and with an average of 145 individuals per image are utilized. Detecting faces in the crowded scenes with the underlying challenges cannot be addressed using a single face detection method. In this paper, by incorporating different traditional machine learning (ML) and the state-of-the-art convolutional neural network (CNN) algorithms, a robust method for counting faces visible in the crowd is proposed. Utilizing a compact variant of the VGG neural network [21], the proposed algorithm outperforms various state-of-the-art algorithms in detecting faces ‘in-the-wild’.