Refining Yolov4 for Vehicle Detection

Real-time vehicle detection is a technology employed in applications like selfdriving cars, traffic camera surveillance. Every year we see better and updated stateof-the-art (SOTA) object detectors, but as those are trained on general-purpose datasets (like MS COCO), we miss out on targeted model improvements for vehicular data. The aim of this paper is to improve the newly released, YOLOv4 detector, specifically, for vehicle tracking applications using some existing methods such as optimising anchor box predictions by using k-means clustering. We also carefully hand-pick and verify some key techniques mentioned in the original paper, to optimise YOLOv4 as per the requirements of our dataset (UA-DETRAC).<br><br>Our fine-tuned model is also compared with the existing models on a number of performance metrics such as - precision, recall, F1 score, mean average precision, and the average IoU. Our experimental results show that the SOTA model which already has real-time object detection capabilities can be further improved for highly targeted use cases. We urge the readers to expand the scope of the paper (and the original model) to other specific situations as well.