Mapping a Semi-Structured Mixed Environment Using a Data-Driven Occupancy Model

Abstract

The issue of identifying available vacant parking spaces in a parking structure proves to be a prevailing challenge. Several methods implemented in parking structures across the nation have been successful in determining the available spaces and alerting drivers towards such spaces. These methods use overhead light waves and under the pavement pressure sensors to measure each parking space. Regardless of their success in providing an occupancy map, the cost of implementing and maintaining such sensors prove to be burdensome. On the other hand, autonomous guidance and navigation technology implemented in autonomous vehicles have been improving rapidly in the past decade. Autonomous vehicles are now capable of scanning, identifying, and avoiding obstacles. Hence, they are theoretically capable of recognizing available parking spaces and performing parking maneuvers. This study employs autonomous vehicles to map public utility spaces, particularly structured and semi-structured parking areas. The goal is to have an accurate and updated occupancy map of the parking structure. Each autonomous vehicle is assigned a specific zone in the parking structure to map and send the data to a map server. Here, a zone refers to a partition in the occupancy map containing the information about the number of available parking spaces. The autonomous vehicles periodically scan the zones assigned by the optimal allocation problem, resulting in a fully updated occupancy map.

This work uses historical data to generate two dynamic models that predicts the time-varying number of available parking spaces. The two models used are a linear dynamic system model and a Gaussian Process Regression (GPR) model. However, the number of spaces must be known for individual zones within the map. Therefore, the probability of human drivers parking in each zone is calculated by utilizing the zone’s “attractiveness score”. The score is a zone-dependent measure that captures its appeal to human drivers based on walking distance, floor location, and other characteristics. Each model’s output (total number of open parking spaces) is divided into individual zones according to the probability of human drivers parking in each zone. The knowledge regarding zone-wise distribution of available spaces is used to formulate an optimal allocation problem that seeks to maximize Value of Information (VoI) such that the information gathered by scanning the zone will return the greatest increment in knowledge related to the garage’s overall occupancy map.

Finally, efficacy of the system developed is illustrated through numerical simulations as percentage of knowledge of the parking occupancy map is shown for four different optimal zone allocation problem. The result of each problem is shown and discussed. The result showed that for better performance while not controlling the traffic in the parking garage, the multi-zone scanning optimal allocation formulation of the problem is the better choice. However, if the parking garage has narrow aisles, then traffic control is important and can be achieved by the chance-constrained optimal allocation formulation of the problem.