Crowd Counting using Single Branch Convolutional Neural Network

Crowd counting has emerged as a key area of study in computer vision and image processing due to the increasing demand for managing and monitoring crowded settings. To extract the scale-relevant features, current neural-network-based techniques employ a multi-network model, which is more complex for optimization and computational waste. To address this, a lightweight single branch convolutional neural network is proposed for single-image crowd counting, significantly reducing computational overhead without compromising model performance. By integrating Gaussian maps during preprocessing, our model gains better generalization and performance. Additionally, Grad-CAM (Gradient-weighted Class Activation Mapping), an explainable AI (XAI) technique, highlights the most influential regions of the image on the crowd counting forecast using to further improve model interpretability. For this task, state-of-the-art Shanghai dataset is used which has two parts SHHA and SHHB. In the SHHA dataset, the model yields MAE and MSE values of 32.87 and 71.77, respectively, while on SHHB, it attains 6.07 and 14.6. Notably, the model contains 57,825 parameters and FLOPs 3.05M which makes it highly efficient and lightweight.