Estimating Traffic Density Using Convolutional Neural Networks Based on Crowd Counting

The quest for accurate traffic density estimation is gaining momentum globally, with Vietnam distinguished by its ranking among the top ten nations for private vehicle usage. Rapid advancements in computer vision, particularly through the development of convolutional neural network (CNN) methodologies, underscore the pressing need to incorporate these techniques into traffic density estimation efforts. In this study, three convolutional neural network (CNN) models—W-Net, UASD-Net (a fusion of U-Net with Adaptive Scenario Discovery), and CSR-Net (Congested Scene Recognition Network)—are employed to quantify and assess traffic density based on images captured in Vietnam. Furthermore, a novel approach for reallocating label points to generate more accurate density maps is proposed. Experimental results on a composite dataset, integrating the TRANCOS, TayDo, and KienGiang datasets, demonstrate promising mean absolute error rates of 3.67, 4.42, and 3.82 for W-Net, UASD-Net, and CSR-Net, respectively.