Low Latency Deep Learning Based Parking Occupancy Detection By Exploiting Structural Similarity

Abstract

Proliferation of vehicle traffic and increasing demands for parking resources in metropolitan areas have induced the frustration of finding parking among motorists. This paper put forward a novel structural similarity (SSIM) decision scheme to realize low latency outdoor parking occupancy detection that serve as an integral part in smart parking framework by providing parking availability information in real time. An SSIM decision module is added on top of the conventional trained Convolutional Neural Network (CNN) classifier to greatly reduce the reaction time in identifying the occupancy status of parking space images extracted from live parking lot camera feeds. The proposed SSIM based parking occupancy detection has been implemented and deployed at an outdoor carpark in Multimedia University, Malaysia to assess its performance in term of detection accuracy and computation time. Assessment results show that the incorporation of SSIM decision module does not deteriorate the accuracy of parking occupancy classification with an overall accuracy of 99%. The computation time of detection application is shortened by more than six times when compared to the pure CNN classification approach when there is only a single instance of parking occupancy change, registering a processing time of 0.45 second when running on a Raspberry Pi 3 single board computer.