Heavy-Vehicle Detection Based on YOLOv4 featuring Data Augmentation and Transfer-Learning Techniques

Abstract

Real-time Vehicle detection is crucial in today's era for our complex interconnected transportation ecosystem built on an advanced technological network of intelligent systems encompassing a wide range of applications such as autonomous vehicles, traffic Surveillance, advanced driver assistance systems, and etcetera. The significance of its application to digital transportation infrastructure embarks upon a distinct framework for heavy-vehicle detection integrated with the YOLOv4 algorithm for real-time detection. In this proposed work, two entities of heavy vehicles such as buses, trucks are considered. The crux of the model, an algorithmic computational mechanism incorporates Mosaic Data augmentation and Transfer-learning techniques that are applied to avoid over-fitting and to improve the optimal speed during training. Subsequently, a fine-tuning YOLOv4 algorithm is implemented for detecting the heavy vehicle. The algorithm is tested for real-time situations in various traffic densities through Computer Vision. Experimental results show that the proposed system achieves higher detection accuracy of 96.54% mAP. More specifically, the performance of the proposed algorithm with the COCO test set and PASCAL VOC 2007 test set demonstrates improvement when compared with other state-of-the-art approaches.