An Approach to Real Time Parking Management using Computer Vision

Abstract

Automating vehicle statistics provides vital information that can be used in predicting the flow of traffic. Object detection based systems that use computer vision have produced drastic improvements in results over a sensor based approach. The methodology proposed in the paper follows an approach to perform this operation in real time and is currently being used in estimating the density of parking spaces, amongst other applications. The paper describes a 4 layer architecture for parking management which involves a HAAR based frame extraction from live video feed followed by a YOLOv2(You Only Look Once) deep neural network approach that supports real time detection of vehicles. The third layer emphasizes on the use of a mechanism that measures the number of vehicles entering a parking space by following the path traced by the centroid which is followed by a number plate recognition system that can retrace mishappenings to their source. The detection system developed using this model has been extensively tested on real time traffic in Bangalore and has generated accuracies close to 95% with video data that has been cross verified manually, making it much more effective than sensor based models.