Abstract
Crowd estimation is a vital component of crowd analysis. It finds many applications in real-worldscenarios, e.g. huge gatherings management like Hajj, sporting and musical events, or political rallies. Automated crowd counting facilitates better and effective management of such events and consequently prevents any undesired situation. This is a very challenging problem in practice since there exists a significant difference in the crowd number in and across different images, varying image resolution, large perspective, severe occlusions, and dense crowd-like cluttered background regions. Current approaches do not handle huge crowd diversity well and thus perform poorly in cases ranging from extreme low to high crowd-density, thus, yielding huge crowd underestimation or overestimation. Also, manual crowd counting proves to be infeasible due to very slow and inaccurate results. To address these major crowd counting issues and challenges, we investigate two different types of input data: uni-modal (image) and multi-modal (image and audio). In the uni-modal setting, we propose and analyze four novel end-to-end crowd counting networks, ranging from multi-scale fusion-based models to uni-scale one-pass and two-pass multitask networks. The multi-scale networks employ the attention mechanism to enhance the model efficacy. On the other hand, the uni-scale models are well-equipped with novel and simple-yet effective patch re-scaling module (PRM) that functions identical but is more lightweight than multi-scale approaches. Experimental evaluation demonstrates that the proposed networks outperform the state-of-the-art in majority cases on four different benchmark datasets with up to 12.6% improvement for the RMSE evaluation metric. The better cross-dataset performance also validates the better generalization ability of our schemes. For the multi-modal input, effective feature-extraction (FE) and strong information fusion between two modalities remain a big challenge. Thus, the multi-modal novel network design focuses on investigating different features fusion techniques amid improving the FE. Based on the comprehensive experimental evaluation, the proposed multi-modal network increases the performance under all standard evaluation criteria with up to 33.8% improvement in comparison to the state-of-the-art. The application of multi-scale uni-modal attention networks also proves more effective in other deep learning domains, as demonstrated successfully on seven different scene-text recognition task datasets with better performance.