Boosting high density crowd counting via density aware Semi-Supervised Learning

Over the last decade, we have witnessed a large interest of the scientific community in developing innovative techniques based on artificial vision for the automated detection and counting of individuals in images and videos. Despite advancements, counting people in highly crowded environments remains an unresolved challenge. In this paper, we propose a real-time algorithm for crowd counting based on point detection, able to handle both sparse and very dense situations, even in the presence of limited training data for high-density scenarios, by a combining curriculum learning strategy with a semi-supervised approach. In fact, during the training procedure, samples are presented to the network starting with simpler concepts (i.e., images with a few people) and gradually advancing to more complex ones (i.e., images with high density). Additionally, to address the challenge of limited labeled data in very crowded environments, the semi-supervised component is selectively activated only in the latest curriculum stages, namely the medium-hard and hard density levels, where unlabeled data become particularly valuable. The experiment, conducted on a test set composed of about 44,000 images belonging to 18 different publicly available datasets, confirms the effectiveness of the proposed approach. Compared with a state-of-the-art baseline point detection algorithm, we observe impressive average improvements of 73% and 70% in terms of Mean Absolute Error (MAE) and Mean Squared Error (MSE), respectively. Interestingly, in high crowded environments (more than 200 people), the impact of the proposed approach is impressive, proving to be the most effective solution; notably, even in less crowded environments (less than 200 people), the curriculum learning approach alone proves sufficient to manage the complexity of the environment, still overcoming the baseline approach.