Accurate characterization of traffic demand and urban logistics flows is vital for mitigating congestion and optimizing infrastructure in regional hubs like Panama City. However, traffic monitoring remains a challenge in many developing countries like Panama due to the high operational costs and manual labor dependencies of traditional monitoring methods. This research addresses this gap by presenting an integrated approach leveraging object detection and classification techniques that can provide accurate traffic flow inference.

While basic vehicle types can be classified with simpler models, scaling to granular categories requires specialized architectures. This paper addresses this challenge by evaluating state-of-the-art deep learning backbones to establish a robust foundation for increasingly granular classification tasks. We conducted a rigorous benchmarking study evaluating six backbone families for classification: EfficientNet V2, MobileNetV4, Swin Transformer V2, ViT, ConvNext V2 and DINO V3. Using a custom dataset of 7 representative vehicle classes, we analyze both quality and footprint metrics for training and inference under linear probing and partial fine-tuning of the last trainable block, aiming to find the most Pareto-optimal model for vehicle classification.

Results demonstrated that while DINOv3 ViT Base achieves peak accuracy under fine-tuning, DINOv3 ViT Small represents the most Pareto-optimal solution, balancing accuracy and computational cost. Furthermore, TensorRT compilation achieves consistent 9× to 10× speedups across most architectures, reaching up to 18x for ViT-Base and EfficientNet V2-Medium. This research successfully validates the viability of transformer-based backbones for real-time deployment in traffic monitoring pipelines, proving that this architecture is viable for scaling to more granular classification.

Keywords

Backbone, Linear Probing, Pareto-optimality, classification and Traffic