Applications of Deep Learning for Vehicle Detection Using Geospatial Data

Abstract

This chapter details initially acquiring an open-source UAV dataset and creating a Google Earth dataset of vehicles, and creating the metadata for these images. Then training a deep learning object detection model, YOLOv4, to generate the best training weight files, having a very high mean average precision (mAP). It is the measure of how precisely the model is detecting the objects specified in the metadata of the validation dataset. The higher this value, the more accurate the model is, with the specific data it has been trained upon. Then deploying the model on various satellite data of certain areas like parking lots, toll booths, and roads, over a certain period of time, to count the number of vehicles in RGB images and from those images calculate factors like maximum capacity of parking lots, average vehicle density of roads, congestion rate in toll booths, length of congestion in toll booths, etc. The model trained on the UAV Dataset at various conditions of weather, daytime, and different resolutions are tested over other UAV Datasets and the trained weights are uploaded to GitHub for future use.