Files
Abstract
The purpose of this project was to create a proof-of-concept vision based automated parking
space monitoring system to monitor the number of empty parking spaces in a defined area. This
project was specifically targeted toward parking lot J at the University of Southern Indiana with
a desired camera placement on the third floor of the Business and Engineering Center. This lot
was selected as it has limited spaces while being the closest lot to several university buildings,
resulting in high vehicle and pedestrian traffic during peak times. The original idea to solve this
proposed by various faculty was to implement a camera and image processing solution in a thirdfloor window of the USI Business and Engineering building as it would be possible to do so with
no added infrastructure. A computer vision model using the opensource program TensorFlow
was selected as a solution to this problem due to the relatively low cost and minimal necessary
infrastructure for operation. This software was implemented on an NVIDIA Jetson Nano using
the SSD Mobilenet detection architecture. A transfer learning approach was used to retrain the
SSD Mobilenet detection architecture to tailor it to this application. Custom datasets were
generated to improve results. An additional detection model was created to automate parking
space location detection. The final design is able to accurately detect vehicles within a
designated area. Due to limitations regarding perspective and image processing resolution, the
final design cannot be implemented solely through camera placement on the third floor of the
Business and Engineering building. A cost analysis determined that the added infrastructure
required to implement a multi-camera solution would be similar to the cost of traditional sensor based solutions.