Model Free Identification of Traffic Conditions Using Unmanned Aerial Vehicles and Deep Learning

The purpose of this paper is to provide a methodological framework to identify traffic conditions based on non-calibrated video recordings captured from unmanned aerial vehicles (UAV) using deep learning. To this end, we propose two complementary to each other approaches: (i) identify in real time, with minimal computational cost, traffic conditions, (ii) localize, classify vehicles and approximate traffic variables (volume, speed, density) on a road segment from video captured by UAVs. Both problems are formulated as classification problems and tackled using Convolutional Neural Networks (CNN). The use of pre-trained CNNs is also investigated. Both approaches are, then, analysed based on their accuracy and feasibility in implementation. Findings indicate that all models developed achieve a detection accuracy of 89% and higher. The CNN with the best performance can classify traffic conditions between constrained and unconstrained traffic with 91% accuracy higher than what a pretrained model achieved and with significantly faster training times. Furthermore, findings indicated that pretrained neural network for traffic localization was able to predict the position and type of vehicles with a precision of 0.91. Based on the fundamental traffic diagram, it was shown that the two approaches provide compatible results and a feasible representation of traffic on the study area. Finally, possible applications in the field of transportation and traffic monitoring are discussed.