Enhancing YOLOv8 for Vehicle Detection in Intelligent Traffic Management

Abstract

Object detection has witnessed significant advancements with the introduction of deep learning models. In this study, we evaluated YOLOv8 model on the BDD100K dataset over 25 epochs, focusing on its progression in precision, recall, mean Average Precision (mAP), and loss metrics. The YOLOv8 model was trained and evaluated on the BDD100K dataset. Data preprocessing was done by resizing images and applying augmentation techniques like scaling, flipping, and cropping to enhance generalization. The model utilized an anchor-free detection head and an optimized backbone network for improved performance. Training was conducted over 25 epochs using the stochastic gradient descent optimizer, with non-maximum suppression and confidence thresholding applied during post-processing to refine detections. Performance was assessed using precision, recall, mAP, and loss metrics. Early epochs were marked by high classification and box losses, which reduced significantly by epoch 5. Precision and recall improved consistently, with precision increasing from 0.81 to 0.74 and recall stabilizing around 0.72–0.78. Middle epochs (6–20) showed stabilization in losses, with box loss reducing to 0.7–0.8 and classification loss to 0.4–0.5. Performance metrics peaked during this phase, achieving a precision of 0.90, recall of 0.84, and mAP@50 and mAP@50-95 values of 0.90 and 0.74, respectively. During the final epochs (21–25), the model achieved optimal stability, with box loss around 0.69 and classification loss at 0.37. Precision remained at 0.90, recall at 0.85, and mAP metrics stabilized at 0.91 and 0.74. A comparative analysis underscored YOLOv8’s superiority over YOLOv5, achieving higher F1-score (0.87 vs. 0.67), mAP@50 (0.91 vs. 0.682), and mAP@50-95 (0.74 vs. 0.449). YOLOv8’s architectural advancements like anchor-free detection head and optimized backbone improved object detection. YOLOv8’s balance of precision and recall, supported by strong post-processing, affirms its robustness for real-world applications.