Vehicular Traffic Flow Parameter Estimation Using Deep Neural Networks

Abstract

To overcome the challenges regarding the fixed place camera installation we introduce a drone based real time traffic flow parameter estimation. With the recent convenience of UAV (Unmanned aerial vehicles) videos begins a new smart transportation application. We Captured the live vehicles traffic data using DJI Phantom 4 pro drone from 8 AM - 6 PM. We manually annotated the vehicles instances of about 6000 using LabelImg tool. Our dataset comprises of 6000 images of resolution 3840 x 2160 and 4096 x 2160 and collectively contain 35000 vehicle instances in 17 different categories of vehicles that are bike, car, bus, suzuki pickup, suzuki bolan, mazda container, Double cabin, Mini Van, Van, Cart, Tractor, Cement Mixer Container, Water Tank, Pajero, Jeep and Coaster. These vehicle instances will be annotated by incorporating there purpose and classes in mind. We manually annotated the speed estimation dataset using on screen pixel measurement tool. We calibrate the frames with the ground truth real frames to generate accurate results related to traffic flow parameter. With the help of our customized vehicles dataset and models, we will be able to detect vehicles of diverse nature of classes (buses, cars, bikes, trucks, mini-trucks, MPVs) and estimate real time traffic flow. Now we are capable to find the traffic flow parameters on local highways. We prepared dataset of different classes and estimate real time traffic flow of vehicles in transportation industry from different road segments on Kashmir HighWay Vehicle Dataset known as KHWD. Our model can be used for vehicles detections, traffic flow estimations, Vehicles speed estimation, anomaly detection, vehicle surveillance, Lane changing behavior, traffic monitoring. Our model will be applicable where it is difficult to install camera sensors or radar sensors.