Parking Occupancy Detection: A Lightweight Deep Neural Network Approach

Abstract

Inaccessibility of real-time parking occupancy information may cause inefficiency in parking management. This paper proposed a novel lightweight deep neural network approach to realize outdoor parking occupancy detection system to support more efficient parking management. A lightweight MobileNet binary classifier is used to accurately identify the occupancy status of parking space image patches that are extracted from live parking lot camera feeds. A performance comparison between different network configurations of MobileNet has been done to investigate their speed-accuracy trade-off when running on embedded device. The prototype was deployed at an outdoor campus parking to evaluate effectiveness of the proposed system. The prototype can detect 22 parking spaces within 2.4 s when running on an ASUS Tinker Board and achieve a detection accuracy of 99%.