Date of Award

6-1-2019

Document Type

Thesis (Undergraduate)

Department

Department of Computer Science

First Advisor

Timothy J. Pierson

Abstract

The issue of parking is more at the forefront of urban development than one might believe. In fact, academic studies have shown that roughly 30% of city traffic is due to drivers circling city blocks attempting to find an open spot. Due to such congestion people often avoid urban centers and downtown areas for shopping or dining because parking is such a hassle and assumed to be unavailable. If drivers knew where parking was available in real time, they could proceed directly to open spaces as opposed to their congestion-inducing attempts to park. A better solution would guide drivers to available parking and may help re-vitalize downtown areas. The problem of knowing whether an available spot exists, however, is complex. This thesis is an investigation and analysis of the efficacy of magnetometers as vehicle sensors for on-street (non-garage) parking. While many solutions to detecting available parking have been tried, we focused on magnetometer-based vehicle sensors placed in each parking spot. We built a sensor comprised of a low-cost magnetometer, a radio, a micro-controller, and a battery on a custom printed circuit board. Our idea is that such a sensor could be placed in each parking space and monitor for vehicles. When a vehicle arrives, the magnetometer detects a change in the magnetic environment, then radios the presence of the vehicle in a space to a central server that aggregates and disseminates parking data to drivers and city officials. City officials could use this data to craft better parking policies and prices. Drivers could then use GPS coordinates and the aggregated space availability in navigation apps to proceed directly to open spaces. Our hope is that this work will provide a foundation for others to learn from our insights into the reliability, stability, and accuracy of such parking sensors. After obtaining permission from Dartmouth Parking and Transportation Services, we conducted experiments in a surface lot on Dartmouth College's campus, and as such, we limited our data collection to a single grid-like parking arrangement to gain deeper insight to one common mode of parking. The analysis of the collected data leverages machine learning via sci-kit learn to form a robust detection algorithm for whether a vehicle is in a space. We utilized four detection algorithms in total, one via a simple magnitude threshold, another using Gaussian Bayes classification, a decision tree classification model, and finally a random forest model. All of these methods succeeded in correctly detecting the status of a parking spot with accuracy well above 90%. Our best classification model, which uses a decision tree, correctly predicted parking space occupancy with 99% accuracy. In our experiments we show that these sensors are stable and do not drift from their initial reading. Our detection algorithms show that they are an accurate option for vehicle detection. Finally, we show that the placement of a sensor is not crucial, so long as the sensor is centrally placed in a parking spot.

Comments

Originally posted in the Dartmouth College Computer Science Technical Report Series, number TR2019-870.

Share

COinS