Rotated Mask Region-Based Convolutional Neural Network Detection for Parking Space Management System

Abstract

Parking space management systems help organize and optimize available parking spaces for consumers, making the process of finding and using parking spaces more efficient. Current parking space management systems include manual recognition, the employment of magnetic and ultrasonic sensors, and, recently, computer vision (CV). One relatively new region-based convolutional neural network (R-CNN) model, Mask R-CNN, has shown promise in its ability to detect objects and has demonstrated superior performance over many other popular CV methods. Building on Mask R-CNN, an updated version, Rotated Mask R-CNN, which can generate bounding boxes the axes of which are rotated with respect to the image’s axis, was proposed to address the limitation of Mask R-CNN. Albeit with the documented theoretical benefits, the application of the rotated version is rare because of its recent invention. To this end, the study aims to detect vehicle instances in one parking lot using various Rotated Mask R-CNN models based on unmanned aircraft system collected images. Both average precision and average recall were utilized to assess the performance of the alternative models with different backbone and head networks. The results reveal the high accuracy level associated with Rotated Mask R-CNN in real-time detection of vehicles. In addition, the results indicate that the inference speed and total loss are highly correlated with head networks and training schedules.