DEEP LEARNING IN VEHICLE DETECTION

Abstract

The present work is based on acquiring open-source UAV dataset and creating Google Earth Dataset of Vehicles and creating the metadata for these images for training of individual Deep Learning Object Detection Models specific to these datasets or similar datasets or data acquired in the similar way, i.e., from flying an UAV or from RGB stacked satellite images which are available in Google Earth. The challenge to train a Deep Learning Object Detection Model, YOLOv4, to generate the best training weight files having a very high mean Average Precision (mAP). This value is the measure of how precisely the model is detecting the objects specified in the metadata of the validation dataset. If this is high that means the more accurate the Deep Learning Model is, with the specific data it has been trained upon. After training the models to an acceptable standard, the goal is to deploy 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 and 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. Thus, the data acquired from Google Earth which is a stitched-out globe of data from various satellites sold to Google by Maxar Technologies, is used for training and testing of the deep learning model which can be then used over any RGB satellite data from Google Earth of any place, at any time and of a certain resolution to calculate these important factors. The deep learning model trained on the open-source UAV Dataset at various conditions of weather, daytime and different resolutions is tested over other UAV Datasets and the trained weights made available in open-source for future use on UAV data and also for future training over advanced UAV datasets with the help of transfer learning to create even more advanced deep learning models.