Boosting real-time vehicle detection in urban traffic using a novel multi-augmentation

Abstract

Real-time vehicle object detection in urban traffic is crucial for modern traffic management systems. This study focuses on improving the accuracy of vehicle identification and classification in heavy traffic during peak hours, with particular emphasis on challenges such as small object sizes and interference from light reflections. The use of multi-label images enables the simultaneous detection of various vehicle types within a single frame, providing more detailed information about traffic conditions. You only look once (YOLO) was chosen for its capability to perform real-time object detection with high accuracy. Multi-augmentation techniques were applied to enrich the training data, making the model more robust to varying lighting conditions, viewpoints, object occlusions, and issues related to small objects. YOLOv8n and YOLOv9t were selected for their speed and efficiency. Models without augmentation, 10 single-augmentation techniques, and 5 multi-augmentation techniques were tested. The results show that YOLOv8n with multiaugmentation (scaling, zoom in, brightness adjustment, color jitter, and noise injection) achieved the highest mAP50-95 score of 0.536, surpassing YOLOv8n with single-augmentation Blur, which had an mAP50-95 of 0.465, as well as YOLOv8n without augmentation, which scored 0.390. Multiaugmentation proved to significantly enhance YOLO’s performance.