Abstract:
Dhaka suffers from severe traffic anomalies, whereby congestion affects not only economic productivity but also quality of life, making it the slowest city in the world because of its severe traffic problems. Despite the widespread impact, existing deep learning studies are limited and leave valuable insights untapped, requiring internet connectivity and offering limited real-time analysis of images or live video feeds. I addressed these challenges in a multi-stage Traffic Congestion Region Detection framework that identifies a congested region with object counting, fusing object detection with color-coded bounding boxes and connected components. This is a junction of state-of-the-art, advanced deep learning models for object detection with Explainable AI techniques that provide explanations of model decisions through detailed visuals, hence improving transparency. Since this research is related to working with one of the world's most congested traffic zones in Dhaka, my methodology consisted of extensive data preparation by doing exploratory data analysis and data augmentation for model generalizability with traffic images. The TCRD framework achieved high detection accuracy with the YOLOv10 model, among both YOLOv9 and YOLOv10, attaining a 94% mAP, 94% precision, and 91% recall score. Employing Eigen-CAM confirmed the accuracy, transparency and potential for improvement in model decisions. In particular, this new multi-stage framework can really identifythe congestion regions of different categories in an image and give insights into other traffic anomalies, hence hugely stepping toward the improvement of traffic management solutions in urban environments like Dhaka. As a result, my solution can be integrated into enhanced traffic management and urban planning strategies to realize more efficient and effective solutions for traffic anomalies.
