Sistema de reconocimiento sobre la disponibilidad de zonas para parqueo mediante redes neuronales convolucionales con imágenes en tiempo real en el Campus Sur de la Universidad Politécnica Salesiana.

Abstract

The objective of this work was to develop an image recognition application about parking spot availability detection, based on convolutional neural networks. The neural network was implemented through mAlexnet architecture due its accuracy and response time the objective data set CNRPark-Ext. Parking zone images are extracted in real time through HIKVISION software, manufacturer of the surveillance cameras. Subsequently images are segmented, preprocessed, classified and stored in a database through algorithms codified in PYTHON. Finally, parallel computing was applied for parking spot classification through the PYTHON MULTIPROCESSING library. Output results from image classification showed a 97.37% accuracy with the dataset CNRPark+Ext and 95.91% with the target dataset. Performance tests with parallel computing allowed us to conclude that this approach only takes advantage from the sequential approach when the processing scenario possess a neural network with a long response time or a high amount of parking spots to process concurrently, this owing to the initialization time needed to begin a parallel execution