ENHANCED VEHICLE DETECTION AND CLASSIFICATION THROUGH OPTIMIZED YOLOV4 USING VISION-BASED TECHNOLOGY

Abstract

This ”research aims to improve the detection and classification of vehicles in intelligent transportation systems. A novel model, called YOLOv4_AF, is proposed as an optimized version of YOLOv4. The model addresses challenges related to quick and accurate vehicle detection and identification, such as small gaps between vehicles and interference in images or video frames. It incorporates an attention mechanism to reduce interference features in both the channel and spatial dimensions of the images. Additionally, a modified version of the Feature Pyramid Network (FPN) from the Path Aggregation Network (PAN) is utilized to enhance the effectiveness of down-sampling and improve object positioning in 3D space. Experimental results on the BIT-Vehicle and UA DETRAC datasets demonstrate the superiority of the proposed YOLOv4_AF model over the original YOLOv4 model, as well as two other state-of-the-art models, Faster R-CNN and EfficientDet. The proposed model achieves impressive mean average precision (mAP) values of 83.45% and 77.08%, along with F1 scores of 0.816 and 0.808, respectively, on the two datasets.”