Smart City Traffic Monitoring:YOLOv7 Transfer Learning Approach for Real-Time Vehicle Detection

Abstract

Real-time vehicle detection is a critical component of traffic monitoring, with significant implications for smart city applications. Accurate and efficient detection of vehicles can improve traffic flow and reduce congestion. This paper presents a real-time vehicle detection system based on Deep Learning (DL) techniques, using the YOLOv7 object detection framework. The system was trained on a novel dataset with a diverse range of vehicles, including different sizes, orientations, and lighting conditions, to improve object detection accuracy. To reduce the required training time and computing resources, Transfer learning is utilized to fine-tune two variants of YOLOv7, YOLOv7-x and YOLOv7-tiny. The results of the experiment revealed that the YOLOv7-x achieved a mean Average Precision (mAP) rate of 96.7%, while the YOLOv7-tiny achieved 89.3%. Furthermore, when the models were fed with a video stream, the YOLOv7-tiny achieved 57 Frames Per Second (FPS), whereas the YOLOv7-x achieved 27 FPS. As a result, the YOLOv7-tiny is more suitable for resource-constrained devices, such as those frequently utilized in IoT applications, due to its smaller model size, lower computational requirements, and higher FPS rate with acceptable accuracy. On the other hand, if higher accuracy is the priority, the YOLOv7-x model should be considered. The proposed frameworks help to improve the effectiveness of traffic management systems, leading to more efficient and sustainable transportation in smart cities.