Improving Vehicle Classification and Detection with Deep Neural Networks

摘要

Vehicle detection and classification are major functions of advanced driver assistance systems (ADAS). In this paper, a deep-learning approach for vehicle detection and classification is discussed and improved. More specifically, we utilized the state of the art object detection model “YOLOV4” with a clear focus on vehicles which made the detection process more robust. We conducted real life tests on vehicle images from Egypt at El-Sahil bridge, Imbaba and El-Hussary where our improved model detects the vehicles reliably. Our approach mainly relied on using ResNet50 and VGG16 as a classification backbones for YOLOv4, we used the GTI dataset to train and fine tune both networks to get the one with the better classification accuracy. The mean average precision (MAP) increased by 6.7% for VGG (84.49% to 91.02%) and 7.35% for ResNet (85.3% to 92.653%) then we trained YOLOv4 using OpenImages and Highway data sets for vehicle detection reaching an improvement of nearly 6.3% MAP (86.5% to 92.82%).