Identifying parking spaces detecting occupancy using vision-based IoT devices

摘要

The increasing number of vehicles in high density, urban areas is leading to significant parking space shortages. While systems have been developed to enable visibility into parking space vacancies for drivers, most rely on costly, dedicated sensor devices that require high installation costs. The proliferation of inexpensive Internet of Things (IoT) devices enables the use of compute platforms with integrated cameras that could be used to monitor parking space occupancy. However, even with camera-captured images, manual specification of parking space locations is required before such devices can be used by drivers after device installation. In this paper, we leverage machine learning techniques to develop a method to dynamically identify parking space topologies based on parked vehicle positions. More specifically, we designed and evaluated an occupation detection model to identify vacant parking spaces. We built a prototype implementation of the whole system using a Raspberry Pi and evaluated it on a real-world urban street near the University of Washington campus. The results show that our clustering-based learning technique coupled with our occupation detection pipeline is able to correctly identify parking spaces and determine occupancy without manual specication of parking space locations with an accuracy of 91%. By dynamically aggregating identied parking spaces from multiple IoT devices using Amazon Cloud Services, we demonstrated how a complete, city-wide parking management system can be quickly deployed at low cost.