YOLOv5-GT: A Balanced Improvement in Object Detection Speed and Accuracy for Autonomous Vehicles in Indonesian Mixed Traffic

Object detection speed and accuracy are critical aspects of the perception system in autonomous vehicles. Speed improvement helps the object detector model achieve performance close to real-time, while accuracy enhancement ensures robust detection across various scenes. Balancing these improvements enhances the safety of autonomous vehicles, particularly in mixed and dense traffic conditions, such as those in Indonesia. This study develops a detection model using local datasets to reflect real-world conditions. It achieved balanced improvements in frames per second (fps) and mean Average Precision (mAP@50–95) through a modified YOLOv5 deep learning model. The key enhancements of the model include the integration of GhostConv and Transformer layers into the YOLOv5 architecture, referred to as YOLOv5s-GT, as well as the application of image augmentations with various scenarios to the training data. Experimental results show that the modified YOLOv5 outperformed the baseline model in both fps and mAP metrics, achieving 82.6 fps and 80.1% mAP@50–95, compared to the baseline performance of 75.8 fps and 77.7% mAP@50–95.