Deep Learning for Non-Conventional Local Vehicle Detection and Classification using YOLOv11

Mendoza P.; Quispe B.; Ticona S.A.; Calsina H. · Proceedings of the 2025 IEEE 32nd International Conference on Electronics, Electrical Engineering and Computing, INTERCON 2025 (2025) · DOI: 10.1109/intercon67304.2025.11244705

Resumen

Accelerated global urban expansion generates significant challenges in transportation, which must be addressed through emerging technologies such as deep learning and artificial intelligence that can contribute to solving these problems and be addressed with the implementation of Intelligent Transportation Systems (ITS). This indicates that vehicle detection and classification systems have improved significantly in recent years. The objective of this study was to automatically detect and classify non-conventional local vehicles using the YOLOv11 deep learning model, focusing on three classes: Combis, Mototoritos, and Triciclos. A dataset of 1496 annotated images was developed and enhanced through data augmentation techniques. The YOLOv11n model trained on this dataset achieved an accuracy of mAP0.50:0.95 0.868 and a mAP0.50 of 0.76, demonstrating competitive performance and highly effective detection for non-conventional vehicle types. It also proved suitable for real-time applications with an average inference time of 125ms per image (∼80 FPS). The contextualized dataset mitigated domain shift and maximized model efficiency. The results confirm that training with a locally adapted dataset significantly reduced domain change, allowing reliable vehicle detection for Intelligent Transport Systems in environments where global datasets are insufficient. This work can be improved by expanding the dataset to include new vehicle classes and integrating the model into embedded platforms.

Instituciones: Universidad Nacional de Juliaca