Prototype of Vehicular Flow Analysis System using Neural Networks

Abstract

This paper presents a prototype of vehicle flow analysis system for the city of Posadas, developed by the Urban Mobility Secretariat of the Municipality of Posadas. The system allows direction and turn detection and vehicle counting based on a video captured by a surveillance camera or a drone. The prototype is based on a YoloV8 neural network for vehicle detection, ByteTrack for tracking, and the VisDrone and COCO datasets. The videos are from an intersection of two heavily congested avenues that serve as entrances to the city center. The system allows for defining entry and exit zones to count vehicles entering the intersection and their directions, thereby identifying the most frequently used turns when vehicles enter the intersection. While similar systems exist, they tend to be very complex and expensive in licensing. The possibility of replicating this solution in other municipalities is very feasible since the entire system has been built with free software and only requires videos captured with any device.