A New Method for Lightweight Extracting Traffic Parameters from Drone Videos and Experimental Verification

Abstract

In order to study the accurate extraction of vehicle traffic parameter information from drone aerial videos to improve urban traffic intelligent management and auxiliary modeling of car-following behavior. A new method for lightweight extraction of vehicle following behavior parameters from drone videos is proposed in this article. The improved ShuffleNet network and GSConv module were introduced into the Yolov7-tiny neural network model as the target detection stage. HOG features and IOU motion metrics are introduced into the DeepSort multi-object tracking algorithm as the tracking matching stage. By building a self-built UAV aerial traffic data set, experiments were conducted to prove that the new method improved a few detection and tracking indicators. In addition, it improves the false detection, missed detection, wrong ID conversion and other phenomena of the previous algorithm, and improves the accuracy and lightweight of multi-target tracking. Finally, the velocity and headway parameters extracted from the car-following behavior using the new method were compared with GPS/INS and proved that the errors were within an acceptable range. The newly proposed traffic flow parameters can be used in traffic flow modeling and simulation to improve the dynamic characteristics and safety of the car-following model, thereby alleviating traffic congestion and improving driving safety.