Search

This project deals with the implementation of a metadata visualisation tool for the neural network developed by Ugiat Technologies. As it is a video processing software, the implemented tool is a player with extra features that allows to check if the metadata provided by the neural network is right. It has been done using web technologies for better portability, mainly in JavaScript. This way almost any device can use it, but it is created for desktop use, no matter the operating system, since it runs in a web browser, but the screen resolution should be the one of a laptop or a desktop computer, no mobile usage is covered by the design. For a better user experience, it allows voice commands to control some basic functions like playing, pausing and stopping the video as well as for enabling and disabling some of the extra features added to the video player. All the implementation is completely open source and it allows to continue adding functionalities. The product obtained is good not just for helping the development and tests of the processing software, but also for making demos of its performance.

Automatic summarization generation of sports video content has been object of great interest for many years. Although semantic descriptions techniques have been proposed, many of the approaches still rely on low-level video descriptors that render quite limited results due to the complexity of the problem and to the low capability of the descriptors to represent semantic content. In this paper, a new approach for automatic highlights summarization generation of soccer videos using audio-visual descriptors is presented. The approach is based on the segmentation of the video sequence into shots that will be further analyzed to determine its relevance and interest. Of special interest in the approach is the use of the audio information that provides additional robustness to the overall performance of the summarization system. For every video shot a set of low and mid level audio-visual descriptors are computed and lately adequately combined in order to obtain different relevance measures based on empirical knowledge rules. The final summary is generated by selecting those shots with highest interest according to the specifications of the user and the results of relevance measures. A variety of results are presented with real soccer video sequences that prove the validity of the approach., Peer Reviewed, Postprint (published version)

Els humans tenim una capacitat innata pel reconeixement d'altres persones a través de la cara independentment de les modificacions que aquesta pugui sofrir ja sigui a través de l'edat o essent camuflada per unes ulleres o una barba. En aquesta tesi es vol intentar desenvolupar un sistema complert de reconeixement facial a través d'un ordinador, es a dir, implementar un sistema autònom capaç de reconèixer persones d'una base de dades a partir d'una imatge desconeguda pel sistema i obtinguda de qualsevol font dins d'Internet (Twitter, Facebook, Google, etc.). La idea doncs és per una banda crear una base de dades (que anomenarem watchlist) a partir d'imatges d'Internet suficientment potent i poder així entrenar a un reconeixedor amb la suficient fiabilitat que ens permeti obtenir un grau alt d’èxit de reconeixement. I d'altra banda, poder ser capaços de reconèixer el major numero d'imatges possible que trobem a Internet. Per assolir el nostre objectiu s'ha implementat un sistema d'adquisició i processat d'imatges que adaptarà les fotografies extretes d'Internet perquè el sistema de reconeixement funcioni correctament. D'aquesta manera no és necessari generar una nova base de dades mitjançant la realització de fotografies específiques a la persona que volem reconèixer, ja que en la majoria dels casos això no ens serà permès. Així doncs, s'ha utilitzat un codi ja implementat pel reconeixement facial aprofitant els recursos que ofereix OpenCV [1] i que ens ajudarà a testejar l’eficiència de la watchlist a la vegada que s'ha implementat un detector de cares i ulls juntament amb una sèrie d'algoritmes que milloren l'eficiència de la detecció capaç de normalitzar una imatge de manera automàtica. El resultat, aplicat a diferents imatges de diferents persones, serà la nostra watchlist i servirà per entrenar el reconeixedor i reconèixer noves cares. L’eficiència d'aquesta watchlist automàtica serà contrastada amb una altra watchlist creada manualment on l'usuari marcarà la pos

Automatic generation of sports highlights from recorded audiovisual content has been object of great interest in recent years. The problem is indeed important in the production of second and third division leagues highlights videos where the quantity of raw material is significant and does not contain manual annotations. Many approaches are mostly based on the analysis of the video and disregard the important information provided by the audio track. In this paper, a new approach that combines audio and video descriptors for automatic soccer highlights generation is proposed. The approach is based on the segmentation of the video contents into shots that are further analyzed in order to determine its relevance and interest. These video-shots are scored taking into account the fusion between different audio and video features. The paper is mainly focused to emphasize the importance of audio detectors that play a key role in the analysis and scoring of the video-shots. Specifically, a new algorithm for referee's whistle detection is proposed. The algorithm has been proven to be very robust and efficiently discriminates professional whistles against other types of noises such as public cheering-up, music instruments, etc. Several results have been produced using real soccer video sequences that prove the validity of the proposed audio and video fusion scheme., Peer Reviewed, Postprint (published version)

IC Reverse engineering is the process to analyze an integrated circuit to obtain information about its design, materials, logic circuitry, functionality, performance and other relevant features. The increasingly complexity of microchips using a greater number of layers and logic gates makes this process unaffordable when using traditional methods that rely on human inspection and analysis. Therefore, digital image processing is presented as a fruitful field for automation. In this paper a system for the circuitry extraction, analysis and presentation is described. It is divided in three blocks: 2D Image Tiling, Logic Gates Localization & Recognition and Microchip Navigator. The paper presents an overview of the complete system and is mainly based on the description of the image processing algorithms that are applied to the different blocks such as image stitching, customized Scale Invariant Feature Transform (SIFT) and logic gate localization & recognition., Peer Reviewed, Postprint (published version)

Entregable E.3.2.1., Peer Reviewed, Postprint (published version)

Deriberable 4.1.1, Preprint

The concept of “smart” cars or intelligent vehicles is presented as one of the most promising solutions to reduce the high mortality rate that occurs on the world’s roads nowadays. Besides, the recent publication of standards as the European Standard for Intelligent Transportation System (ITS) or the international standard IEEE 802.11p confirm the importance of the future vehicle-to-vehicle or vehicle-to-infrastructure networks, which can diminish gridlocks or aid the driver with information about the road status or the weather forecast in order to prevent accidents, for instance. The main drawbacks of such intelligent network regarding the accident prevention lie in the presence of obstructing objects on the road or cars that do not implement this V2V system. Therefore, a radar application based on the used waveform for the V2V communication can be suitable as a direct method to avoid collisions. The aim of this thesis consists in the verification of the viability of a radar application in a V2V scenario. Thus, a thorough evaluation of the implemented V2V propagation channel has been performed so as to determine the main constraint factors, such as the power fading, Non Wide Sense Stationary Uncorrelated Scattering (NWSSUS) or the Doppler frequency, that can affect at detection and location application. From the conclusions obtained based on the characterization of a V2V channel, a radar algorithm has been designed as well as a tracking system. The design of the proposed radar algorithm is based on power peak detection that the estimated channel impulse response presents based on the reflected power originated by possible targets. The trilateration method is used for the location of these targets in the azimuth plane; thus, a Multiple Input Multiple Output system is required. In order to carry out this viability study, a MIMO structure 4x4 using OFDM with PSK or QAM as the modulation and over a real V2V propagation channel has been simulated. The Geometric Stochast

OFDM se presenta como una modulación para la nueva generación de comunicaciones móviles (4G), la cual ha sido empleada en estándares como DAB / DVB, wireless LAN, xDSL, WiMAX y PLC; siendo una de las modulaciones multiportadora más populares gracias a las ventajas obtenidas por el uso de varias subportadoras ortogonales entre ellas. Sin embargo OFDM presenta ciertos inconvenientes como el elevado valor de PAPR que se obtiene en el dominio temporal, el cual afecta a los amplificadores de potencia no lineales introduciendo distorsiones. El objetivo principal de este trabajo es la búsqueda, la implementación la adaptación y la evaluación de técnicas de reducción de PAPR para así paliar en parte los efectos de distorsión de los amplificadores no lineales. Para ello, se ha analizado la modulación para simular en Matlab un sistema de comunicación basado en OFDM y en el estándar IEEE 802.11a, en el cual se han introducido las técnicas para evaluar sus rendimientos. El trabajo se ha estructurado en 3 capítulos. En el primero de ellos se analiza la modulación, se explica el concepto de PAPR y las características del sistema simulado. En el siguiente capítulo se introducen las técnicas de reducción de PAPR y se detalla cómo han sido modificadas y adaptadas para el sistema OFDM implementado en Matlab. Mientras que en el último capítulo y en los anexos se evalúa el rendimiento de estas técnicas. El resultado final del trabajo se concluye con una valoración positiva en cuanto a la reducción de la PAPR y la mejora de la BER, a pesar de haber sacrificado básicamente tiempo de procesado debido a la complejidad de las técnicas simuladas.