Efficient face detection based crowd density estimation using convolutional neural networks and an improved sliding window strategy

Counting and detecting occluded faces in a crowd is a challenging task in computer vision. In this paper, we propose a new approach to face detection-based crowd estimation under significant occlusion and head posture variations. Most state-of-the-art face detectors cannot detect excessively occluded faces. To address the problem, an improved approach to training various detectors is described. To obtain a reasonable evaluation of our solution, we trained and tested the model on our substantially occluded data set. The dataset contains images with up to 90 degrees out-of-plane rotation and faces with 25%, 50%, and 75% occlusion levels. In this study, we trained the proposed model on 48,000 images obtained from our dataset consisting of 19 crowd scenes. To evaluate the model, we used 109 images with face counts ranging from 21 to 905 and with an average of 145 individuals per image. Detecting faces in crowded scenes with the underlying challenges cannot be addressed using a single face detection method. Therefore, a robust method for counting visible faces in a crowd is proposed by combining different traditional machine learning and convolutional neural network algorithms. Utilizing a network based on the VGGNet architecture, the proposed algorithm outperforms various state-of-the-art algorithms in detecting faces ‘in-the-wild’. In addition, the performance of the proposed approach is evaluated on publicly available datasets containing in-plane/out-of-plane rotation images as well as images with various lighting changes. The proposed approach achieved similar or higher accuracy.