Image recognition in smart cities using neuronal networks

Abstract

L'objectiu principal és desenvolupar un servei de detecció d'incidències en l'entorn urbà que es troben en mal estat on els ciutadans puguin denunciar-ho penjant una fotografia del lloc on es troba. La fotografia serà analitzada per una xarxa neuronal mitjançant la tècnica de detecció d'objectes per identificar si hi ha un incident i de quin tipus, i si n'hi ha, la ubicació serà informada en una plataforma centralitzada.

El objetivo principal es desarrollar un servicio de detección de incidencias en el entorno urbano en mal estado donde los ciudadanos puedan denunciar subiendo una foto del lugar donde se encuentra. La foto será analizada por una red neuronal mediante técnica de detección de objetos para identificar si hay algún incidente y de qué tipo, y si lo hay, se determinará la ubicación. informado en una plataforma centralizada.

As a consequence of climate change, the search for sustainable development and the improvement of citizens' quality of life, in recent years much has been invested in bringing cities closer to the concept of smart city. Technologies have been developed to manage lighting control in the city, waste collection or, for example, traffic management. The main objective is to develop a service to detect incidents in the urban environment that are in poor condition where citizens can report by uploading a photo of the place where it is located. The photo will be analysed by a neuronal network using object detection technique to identify if there is an incident and of what type, and if there is, the location will be reported on a centralised platform. Images from three different class labels are analysed: broken stop sign, crosswalk in poor condition and fallen trees. In order to correctly detect the class labels in the analysed images, the neuronal network has been trained with different weights to find the most optimal analysis percentage using supervised training. To choose which image to analyse, a platform form is used to choose which class label to search for in the images sent by citizens and a map to show the locations of the damaged urban environment. This request is sent via an API, which sends the command to the neuronal network to analyse the images. Once the results are obtained, it stores the images together with the type of incident and the location in a database located in the cloud, which will be consulted by the centralised platform to display the locations on the map. In summary, this project could be applied to other types of needs such as finding lost pets and lost objects, reporting more types of damaged street furniture, reporting the state of health of the city's trees or finding out if we have enough bike lanes in the city by analysing the number of bikes in the city. However, this project requires a significant amount of expertise in neuronal networks, computational resources for training the network and powerful devices to run the network in real-time.