Traffic congestion estimation has long been a critical aspect in intelligent transportation systems and the need for such an automatic and precise congestion estimation system is increasingly demanding. However, existing approaches often overlook the unique traffic patterns in motorbike-centric cities like Ho Chi Minh City, Vietnam. To address this gap, we propose a novel traffic flow estimation framework, combining vehicle detection and an enhanced road segmentation workflow. The proposed system integrates a lightweight instance segmentation module with squeeze-and-excitation and depthwise-separable convolution blocks, enabling automatic mapping of region-of-interest (ROI) with minimal computational cost. This segmentation component achieves 75.4% Dice Score and 73.9% mIoU. For vehicle detection and tracking, we incorporate an object localization framework augmented through custom training on motorbike-rich scenes, achieving 83.1% mAP@50 and delivering consistently high F1-scores across all congestion levels. We also introduce a new benchmark dataset designed to capture motorbike-rich traffic scenes. This framework offers a balanced trade-off between accuracy and efficiency, enhancing congestion analysis in dense, dynamic urban environments.