This project aims to explore new research problems that can be solved by using image and video data analysis. This project will be developed in collaboration with the Lisbon City Council (CML) and addresses challenges that CML launched for the scientific community. The area topic to be addressed is computer vision for knowledge extraction based on aerial and street-level city imagery. We propose a new multi-spatial scale urban fabric dataset and a novel convolutional neural network solution for urban fabric classification tasks. Considering the challenges launched by CML and the already established collaboration between Iscte and CML, this project proposes to explore solutions to a set of urban classification problems by using video and image analysis. In this topic the goal is to identify features of the city shape, namely: i) to estimate the existence of unregistered greenspaces, as green rooftops (since CML is not currently aware of most green rooftops existing in Lisbon) and green back yards (not registered in the city council); ii) to estimate the height and deployment shape and area of Lisbon’s buildings, which will enable to develop an automatic and constantly updated tri-dimensional urban fabric map for Lisbon; iii) to estimate and automatically classify the conservation state of buildings’ facades; iv) among others. Such a methodology will be initially based on the set of aerial and street-level images obtained from CML (during the period of the scholarship). We are also considering exploiting existing city plans freely available at OpenStreetMaps to boost the convolutional neural network's training process, as well as to improve its inference performance. In a second phase, we will explore the use of other types of images such as from google street viewer and the ones directly collected by cameras in drones or city council vehicles (e.g., garbage trucks). For this second phase, we envision submitting a proposal to competitive funding and propose this topic...

Esta operação visa a melhoria do serviço da Polícia Judiciária (PJ) ao cidadão através da automatização do processo operacional de catalogação de imagens fotográficas e comparação com o banco de imagens existentes com dezenas de milhar de imagens, implementando um protótipo que use as mais recentes aproximações da Inteligência Artificial (IA). Em particular, será criada uma solução inovadora para a área da balística, identificada pela PJ como área prioritária, que permitirá identificar com elevada taxa de sucesso e rapidez a correspondência entre projétil e possível tipo de arma utilizado, reduzindo o esforço e tempo de intervenção dos recursos humanos especializados o que terá impacto no processo de produção de prova e consequentemente tornará Portugal um país mais justo. A solução a desenvolver no ISCTE consiste na implementação de um modelo de classificação específico, baseado em redes neuronais siamesas, que permite a identificação de modelos de armas a partir de imagens de invólucros dos projéteis deflagrados (sistema de classificação balística). O treino e teste do modelo fará uso de uma base de dados de imagens, previamente criada, contendo uma grande quantidade de imagens de invólucros de projéteis deflagrados.