VEHICLE ANALYTICS FOR URBAN ROAD POLICY-MAKING USING STATE-OF-THE-ART DEEP LEARNING NETWORKS

Abstract

Vehicle counting is very essential for the urban policymaker especially for traffic signal management, infrastructure development, and assessing the need of various road facilities. Vehicle counting through Deep learning on road images is very promising and can provide various others information like traffic categories, speed, etc. We have used two state-of-the-art deep learning-based vehicle detection frameworks including YOLOv4 and Faster-RCNN with the use of ResBlock, we have modified the original YOLOv4 to get real- time vehicle count from traffic stream, while Inception-ResNet- v2 is used as the backbone of the Faster-RCNN to get the highly accurate count but offline. Both models are trained on the UA- DETRAC dataset. Modified-YOLOv4 has achieved around 89% accuracy, while Faster-RCNN has around 95%.