Parking Space Detection Using a Machine Learning-Enhanced Unmanned Aerial Vehicle in a Virtual Environment

ABSTRACT: Unmanned aerial vehicles (UAVs) have increased in popularity for several diverse applications over the past few years. Parking, especially in crowded parking lots, can be very time-consuming, as a driver must manually search for vacant spaces among many occupied ones. In this work, reinforcement learning—a category of machine learning in which an agent receives inputs from the environment while outputting actions in order to maximize reward—was utilized in tandem with AirSim, a drone simulator developed by Microsoft, to automate a virtual UAV’s movement. A convolutional neural network (CNN) was then utilized to detect both vacant and filled parking spots, which achieved 98% recall and 93% accuracy. Unreal Engine was used to create a custom environment that resembled a parking lot, and the virtual drone was trained using a Deep Q-Network (DQN). The DQN achieved a mean reward of 394.5 in training and 460.4 in evaluation. A pre-trained CNN integrated with the DQN enables the real-time classification of vacant/occupied parking spaces from drone imagery. Results validate the effectiveness of combining reinforcement learning navigation with CNN image classification, demonstrating deployment-ready performance for real-world congested parking applications.