Deep Learning-based Models with YOLOv7 and Convolutional Neural Networks for Vehicle Detection and Recognition

The application of artificial intelligence (AI) technology has become prevalent across various sectors, including transportation and smart city. A key implementation of AI in this domain is traffic monitoring, often relying on license plate recognition to identify vehicles. However, this approach faces limitations when plates are obscured. To address this issue, this research explores a broader approach by recognizing general vehicle attributes, ensuring more accurate identification and comprehensive traffic statistics. The proposed solution integrates the You Only Look Once (YOLO) object detection algorithm and convolutional neural networks (CNN) pretrained models for vehicle attributes recognition. This study utilizes multiple datasets, including Roboflow Vehicle, Stanford Cars, VehicleID, and VCoR, to ensure comprehensive model evaluation. Experimental results indicate that YOLOv7 achieved a mean average precision (mAP) score of 86.1% for vehicle detection, with an average precision (AP) score of 91.5% for the car class. For vehicle make and model recognition, the lightweight EfficientNetV2S model demonstrated the highest accuracy score, achieving 89.8% and 99.2% on the Stanford Cars and VehicleID dataset, respectively. For vehicle color recognition, DenseNet201 models achieved the highest accuracy score of 87% on the VCoR dataset. These findings underscore the effectiveness of integrating YOLOv7 and CNN models for robust vehicle detection and recognition. This research provides a practical solution to the limitations of traditional license plate recognition methods, contributing to the development of more accurate and efficient traffic monitoring systems. Future studies may further optimize the framework for real-time applications and diverse traffic scenarios.