DepthPark : Smart, Cost‐Effective Vision‐Based Indoor Parking Management System Using Single Monocular Depth Estimation

ABSTRACT Urban areas increasingly face challenges due to traffic congestion and limited parking availability. Traditionally designed for outdoor spaces, parking management systems now spearhead innovative endeavors to address the unique challenges of indoor environments. Implementing smart indoor parking solutions in enclosed lots poses significant challenges, as it requires the installation of sensors in every parking space, a costly endeavor, especially in large and older facilities. To alleviate these costs, we propose harnessing visual sensor networks to develop DepthPark: a novel, cost‐effective, computer vision‐based parking management system that uses single monocular depth estimation for real‐time indoor parking management. This system strategically positions cameras to capture license plate data during vehicle entry and exit. Additionally, it orchestrates the movement of mobile cameras on linear guides through a customized deep neural network, to ensure precise and efficient monitoring of expansive parking lots. This setup not only lowers implementation costs but can also reduce bandwidth requirements by minimizing the unnecessary transmission of frames to the server. Mobile cameras capture and log events only when they occur in the lane, and these are then processed by the server computer. The system uses simple monocular cameras, with distance measurement handled by a depth estimation‐based CNN solution to estimate vehicle positions. DepthPark currently identifies two types of parking violations: occupying an unauthorized parking slot and parking across two spaces. Physical implementation of the system on an Intel Core i7‐10750H processor with 16 GB RAM demonstrated high accuracy, including license number recognition accuracy of 98.31% and parking slot classification accuracy of 96%, with a throughput of 30 fps, underscoring the system's efficiency and effectiveness.