Reports de recerca
http://hdl.handle.net/2117/3754
Fri, 19 Jan 2018 09:59:34 GMT2018-01-19T09:59:34ZTowards SLAM with an events-based camera
http://hdl.handle.net/2117/110514
Towards SLAM with an events-based camera
Clavera Gilaberte, Ignasi; Solà Ortega, Joan; Andrade-Cetto, Juan
Event-based cameras have an incredible potential in real-time and real-world robotics. They would enable more efficient algorithms in applications where high demanding requirements, such as rapid dynamic motion and high dynamic range, make standard cameras run into problems rapid dynamic motion and high dynamic range. While traditional cameras are based in the frame-base paradigm - a shutter captures a certain amount of pictures per second -, the bio-inspired event cameras have pixels that respond independently to the change of log-intensity generating asynchronous events. An special appeal for this type of cameras is their low band-width, since the stream of events contain all the information getting rid of the redundancy. This sensors that mimic some properties of the human retina has microseconds latency and 120 dB dynamic range (in contrast to the 60 dB of the standard cameras).
However, the current impact of the event cameras has been tiny due to the necessity of completely new algorithm, there is no global measurement of the intensity which would allow the use of current methods. The fact that an event corresponds to an asynchronous local intensity difference turns out to be a challenging problem if one wants to recover the motion as well as the scene. This article tries to illustrate the several problems that are needed to face when dealing with this problem and some of the different approaches taken.
First of all, we will explain the generative model of the event camera and the preliminaries, followed by the different approaches. Finally will the conclusions and a glossary of the code.
Mon, 13 Nov 2017 17:25:37 GMThttp://hdl.handle.net/2117/1105142017-11-13T17:25:37ZClavera Gilaberte, IgnasiSolà Ortega, JoanAndrade-Cetto, JuanEvent-based cameras have an incredible potential in real-time and real-world robotics. They would enable more efficient algorithms in applications where high demanding requirements, such as rapid dynamic motion and high dynamic range, make standard cameras run into problems rapid dynamic motion and high dynamic range. While traditional cameras are based in the frame-base paradigm - a shutter captures a certain amount of pictures per second -, the bio-inspired event cameras have pixels that respond independently to the change of log-intensity generating asynchronous events. An special appeal for this type of cameras is their low band-width, since the stream of events contain all the information getting rid of the redundancy. This sensors that mimic some properties of the human retina has microseconds latency and 120 dB dynamic range (in contrast to the 60 dB of the standard cameras).
However, the current impact of the event cameras has been tiny due to the necessity of completely new algorithm, there is no global measurement of the intensity which would allow the use of current methods. The fact that an event corresponds to an asynchronous local intensity difference turns out to be a challenging problem if one wants to recover the motion as well as the scene. This article tries to illustrate the several problems that are needed to face when dealing with this problem and some of the different approaches taken.
First of all, we will explain the generative model of the event camera and the preliminaries, followed by the different approaches. Finally will the conclusions and a glossary of the code.Comparing error minimized extreme learning machines and support vector sequential feed-forward neural networks
http://hdl.handle.net/2117/88076
Comparing error minimized extreme learning machines and support vector sequential feed-forward neural networks
Romero Merino, Enrique; Alquézar Mancho, René
Recently, error minimized extreme learning machines (EM-ELMs) have been proposed as a simple and efficient approach to build single-hidden-layer feed-forward networks (SLFNs) sequentially. They add random hidden nodes one by one (or group by group) and update the output weights incrementally to minimize the sum-of-squares error in the training set. Other very similar methods that also construct SLFNs sequentially had been reported earlier with the main difference that their hidden-layer weights are a subset of the data instead of being random. By analogy with the concept of support vectors original of support vector machines (SVMs), these approaches can be referred to as support vector sequential feed-forward neural networks (SV-SFNNs), and they are a particular case of the Sequential Approximation with Optimal Coefficients and Interacting Frequencies (SAOCIF) method. In this paper, it is firstly shown that EM-ELMs can also be cast as a particular case of SAOCIF. In particular, EM-ELMs can easily be extended to test some number of random candidates at each step and select the best of them, as SAOCIF does. Moreover, it is demonstrated that the cost of the calculation of the optimal output-layer weights in the originally proposed EM-ELMs can be improved if it is replaced by the one included in SAOCIF. Secondly, we present the results of an experimental study on 10 benchmark classification and 10 benchmark regression data sets, comparing EM-ELMs and SV-SFNNs, that was carried out under the same conditions for the two models. Although both models have the same (efficient) computational cost, a statistically significant improvement in generalization performance of SV-SFNNs vs. EM-ELMs was found in 12 out of the 20 benchmark problems.
Thu, 16 Jun 2016 08:38:58 GMThttp://hdl.handle.net/2117/880762016-06-16T08:38:58ZRomero Merino, EnriqueAlquézar Mancho, RenéRecently, error minimized extreme learning machines (EM-ELMs) have been proposed as a simple and efficient approach to build single-hidden-layer feed-forward networks (SLFNs) sequentially. They add random hidden nodes one by one (or group by group) and update the output weights incrementally to minimize the sum-of-squares error in the training set. Other very similar methods that also construct SLFNs sequentially had been reported earlier with the main difference that their hidden-layer weights are a subset of the data instead of being random. By analogy with the concept of support vectors original of support vector machines (SVMs), these approaches can be referred to as support vector sequential feed-forward neural networks (SV-SFNNs), and they are a particular case of the Sequential Approximation with Optimal Coefficients and Interacting Frequencies (SAOCIF) method. In this paper, it is firstly shown that EM-ELMs can also be cast as a particular case of SAOCIF. In particular, EM-ELMs can easily be extended to test some number of random candidates at each step and select the best of them, as SAOCIF does. Moreover, it is demonstrated that the cost of the calculation of the optimal output-layer weights in the originally proposed EM-ELMs can be improved if it is replaced by the one included in SAOCIF. Secondly, we present the results of an experimental study on 10 benchmark classification and 10 benchmark regression data sets, comparing EM-ELMs and SV-SFNNs, that was carried out under the same conditions for the two models. Although both models have the same (efficient) computational cost, a statistically significant improvement in generalization performance of SV-SFNNs vs. EM-ELMs was found in 12 out of the 20 benchmark problems.Augmented regular expressions: a formalism to describe, recognize, and learn a class of context-sensitive languages
http://hdl.handle.net/2117/82163
Augmented regular expressions: a formalism to describe, recognize, and learn a class of context-sensitive languages
Alquézar Mancho, René; Sanfeliu, A.
In order to extend the potential of application of the syntactic approach to pattern recognition, the efficient use of models capable of describing context-sensitive structural relationships is needed. Moreover, the ability to learn such models from examples is interesting to automate as much as possible the development of applications. In this paper, a new formalism that permits to describe a non-trivial class of context-sensitive languages, the Augmented Regular Expressions (AREs), is introduced. AREs augment the descriptive power of regular expressions by including a set of constraints that involve the number of instances of the operands of the star operations in each string of the language. Likewise, algorithms are given to infer AREs from string examples and to recognize language strings by AREs. The method for learning AREs consists of a regular grammatical inference step, aimed at obtaining a regular superset of the target language, followed by a constraint induction process, which reduces the extension of the inferred language transforming it into a context-sensitive one. Hence, this two-step approach avoids the difficulty of learning context-sensitive grammars directly from the data. The method for recognizing language strings is also splitted in two stages: matching the underlying regular expression and checking that the resulting star instances satisfy the constraints.
Wed, 27 Jan 2016 18:19:24 GMThttp://hdl.handle.net/2117/821632016-01-27T18:19:24ZAlquézar Mancho, RenéSanfeliu, A.In order to extend the potential of application of the syntactic approach to pattern recognition, the efficient use of models capable of describing context-sensitive structural relationships is needed. Moreover, the ability to learn such models from examples is interesting to automate as much as possible the development of applications. In this paper, a new formalism that permits to describe a non-trivial class of context-sensitive languages, the Augmented Regular Expressions (AREs), is introduced. AREs augment the descriptive power of regular expressions by including a set of constraints that involve the number of instances of the operands of the star operations in each string of the language. Likewise, algorithms are given to infer AREs from string examples and to recognize language strings by AREs. The method for learning AREs consists of a regular grammatical inference step, aimed at obtaining a regular superset of the target language, followed by a constraint induction process, which reduces the extension of the inferred language transforming it into a context-sensitive one. Hence, this two-step approach avoids the difficulty of learning context-sensitive grammars directly from the data. The method for recognizing language strings is also splitted in two stages: matching the underlying regular expression and checking that the resulting star instances satisfy the constraints.Competitive function approximation for reinforcement learning
http://hdl.handle.net/2117/28454
Competitive function approximation for reinforcement learning
Agostini, Alejandro Gabriel; Celaya Llover, Enric
The application of reinforcement learning to problems with continuous domains requires representing the value function by means of function approximation. We identify two aspects of reinforcement learning that make the function approximation process hard: non-stationarity of the target function and biased sampling. Non-stationarity is the result of the bootstrapping nature of dynamic programming where the value function is estimated using its current approximation. Biased sampling occurs when some regions of the state space are visited too often, causing a reiterated updating with similar values which fade out the occasional updates of infrequently sampled regions.
We propose a competitive approach for function approximation where many different local approximators are available at a given input and the one with expectedly best approximation is selected by means of a relevance function. The local nature of the approximators allows their fast adaptation to non-stationary changes and mitigates the biased sampling problem. The coexistence of multiple approximators updated and tried in parallel permits obtaining a good estimation much faster than would be possible with a single approximator. Experiments in different benchmark problems show that the competitive strategy provides a faster and more stable learning than non-competitive approaches.
Mon, 29 Jun 2015 18:57:10 GMThttp://hdl.handle.net/2117/284542015-06-29T18:57:10ZAgostini, Alejandro GabrielCelaya Llover, EnricThe application of reinforcement learning to problems with continuous domains requires representing the value function by means of function approximation. We identify two aspects of reinforcement learning that make the function approximation process hard: non-stationarity of the target function and biased sampling. Non-stationarity is the result of the bootstrapping nature of dynamic programming where the value function is estimated using its current approximation. Biased sampling occurs when some regions of the state space are visited too often, causing a reiterated updating with similar values which fade out the occasional updates of infrequently sampled regions.
We propose a competitive approach for function approximation where many different local approximators are available at a given input and the one with expectedly best approximation is selected by means of a relevance function. The local nature of the approximators allows their fast adaptation to non-stationary changes and mitigates the biased sampling problem. The coexistence of multiple approximators updated and tried in parallel permits obtaining a good estimation much faster than would be possible with a single approximator. Experiments in different benchmark problems show that the competitive strategy provides a faster and more stable learning than non-competitive approaches.Modelling of tidal power with EasyJava simulations
http://hdl.handle.net/2117/23235
Modelling of tidal power with EasyJava simulations
Baranger, Pierre; Grau Saldes, Antoni; Bolea Monte, Yolanda
This report is devoted to the study of a mathematical model to simulte the behaviour of a power generation plant from the ocean tidals. The simulation has been implemented with EasyJava Simulations (EJS) environment.
Recerca d'un model de generació elèctrica a partir de les marees oceàniques
Tue, 17 Jun 2014 07:38:44 GMThttp://hdl.handle.net/2117/232352014-06-17T07:38:44ZBaranger, PierreGrau Saldes, AntoniBolea Monte, YolandaThis report is devoted to the study of a mathematical model to simulte the behaviour of a power generation plant from the ocean tidals. The simulation has been implemented with EasyJava Simulations (EJS) environment.Etude d'un canal d'irrigation MIMO
http://hdl.handle.net/2117/16938
Etude d'un canal d'irrigation MIMO
Chefdor, Nicolas; Bolea Monte, Yolanda; Grau Saldes, Antoni
Treball de recerca fet per l'alumne Nicolas Chefdor durant la seva estada al laboratori de recerca VIS-ESAII a la UPC. Dirigit pels Dr. Yolanda Bolea i Dr. Antoni Grau. Estada del 1/06/2011-30/09/2011
Fri, 16 Nov 2012 11:44:48 GMThttp://hdl.handle.net/2117/169382012-11-16T11:44:48ZChefdor, NicolasBolea Monte, YolandaGrau Saldes, AntoniTreball de recerca fet per l'alumne Nicolas Chefdor durant la seva estada al laboratori de recerca VIS-ESAII a la UPC. Dirigit pels Dr. Yolanda Bolea i Dr. Antoni Grau. Estada del 1/06/2011-30/09/2011Stochastic approximations of average values using proportions of samples
http://hdl.handle.net/2117/14112
Stochastic approximations of average values using proportions of samples
Agostini, Alejandro Gabriel; Celaya Llover, Enric
In this work we explain how the stochastic approximation of the average of a random variable is carried out when the observations used in the updates consist in proportion of samples rather than complete
samples.
IRI Technical Report
Tue, 29 Nov 2011 14:54:00 GMThttp://hdl.handle.net/2117/141122011-11-29T14:54:00ZAgostini, Alejandro GabrielCelaya Llover, EnricIn this work we explain how the stochastic approximation of the average of a random variable is carried out when the observations used in the updates consist in proportion of samples rather than complete
samples.Registration of 3d point clouds for urban robot mapping
http://hdl.handle.net/2117/13936
Registration of 3d point clouds for urban robot mapping
Teniente Avilés, Ernesto; Andrade-Cetto, Juan
We consider the task of mapping pedestrian urban areas for a robotic guidance and surveillance application. This mapping is performed by registering three-dimensional laser range scans acquired with two different robots.
To solve this task we will use the Iterative Closes Point (ICP) algorithm proposed in [8],
but for the minimization step we will use the metric proposed by Biota et al. [10] trying to get advantage of the compensation between translation and rotation they mention. To reduce computational cost in the original ICP during matching, the correspondences search is done with the library Approximate Nearest Neighbor (ANN). Finally we propose a hierarchical new
correspondence search strategy, using a point-to-plane strategy at the highest level and the point-to-point metric at finer levels. At the highest level the adjust error between a plane and it’s n adjacent points describing the plane is computed, if this error is bigger than a threshold then we change the level.
Wed, 16 Nov 2011 13:18:24 GMThttp://hdl.handle.net/2117/139362011-11-16T13:18:24ZTeniente Avilés, ErnestoAndrade-Cetto, JuanWe consider the task of mapping pedestrian urban areas for a robotic guidance and surveillance application. This mapping is performed by registering three-dimensional laser range scans acquired with two different robots.
To solve this task we will use the Iterative Closes Point (ICP) algorithm proposed in [8],
but for the minimization step we will use the metric proposed by Biota et al. [10] trying to get advantage of the compensation between translation and rotation they mention. To reduce computational cost in the original ICP during matching, the correspondences search is done with the library Approximate Nearest Neighbor (ANN). Finally we propose a hierarchical new
correspondence search strategy, using a point-to-plane strategy at the highest level and the point-to-point metric at finer levels. At the highest level the adjust error between a plane and it’s n adjacent points describing the plane is computed, if this error is bigger than a threshold then we change the level.Path planning with pose SLAM
http://hdl.handle.net/2117/12449
Path planning with pose SLAM
Valencia Carreño, Rafael; Andrade-Cetto, Juan; Porta Pleite, Josep Maria
The probabilistic belief networks that result from standard feature-based simultaneous localization and map building (SLAM) approaches cannot be directly used to plan trajectories. The reason is that they
produce a sparse graph of landmark estimates and their probabilistic relations, which is of little value to find collision free paths for navigation. In contrast, we argue in this paper that Pose SLAM graphs can be directly used as belief roadmaps (BRMs). The original BRM algorithm assumes a known model of the environment from which probabilistic sampling generates a roadmap. In our work, the roadmap is built on-line by the Pose SLAM algorithm. The result is a hybrid BRM-Pose SLAM method that devises optimal navigation strategies on-line by searching for the path with lowest accumulated uncertainty for the robot pose. The method is validated over synthetic data and standard SLAM datasets.
Tue, 03 May 2011 09:44:15 GMThttp://hdl.handle.net/2117/124492011-05-03T09:44:15ZValencia Carreño, RafaelAndrade-Cetto, JuanPorta Pleite, Josep MariaThe probabilistic belief networks that result from standard feature-based simultaneous localization and map building (SLAM) approaches cannot be directly used to plan trajectories. The reason is that they
produce a sparse graph of landmark estimates and their probabilistic relations, which is of little value to find collision free paths for navigation. In contrast, we argue in this paper that Pose SLAM graphs can be directly used as belief roadmaps (BRMs). The original BRM algorithm assumes a known model of the environment from which probabilistic sampling generates a roadmap. In our work, the roadmap is built on-line by the Pose SLAM algorithm. The result is a hybrid BRM-Pose SLAM method that devises optimal navigation strategies on-line by searching for the path with lowest accumulated uncertainty for the robot pose. The method is validated over synthetic data and standard SLAM datasets.Estudi de la transformació de l'espai de color RGB a l'espai de color HSV
http://hdl.handle.net/2117/12013
Estudi de la transformació de l'espai de color RGB a l'espai de color HSV
Grau Gotés, Mª Ángela; Grau Sánchez, Miguel; Montseny Masip, Eduard; Sobrevilla Frisón, Pilar
S’apliquen les tècniques clàssiques de propagació de l’error a la transformació de l’espai de color RGB en l’espai de color HSV a un conjunt de 1098 imatges test. El conjunt d’imatges test són 183 paletes de color i sis nivells d’il·luminació diferents. Els resultats que es presenten indiquen com varien la mitjana i la variància per la transformació.
Tue, 22 Mar 2011 10:51:28 GMThttp://hdl.handle.net/2117/120132011-03-22T10:51:28ZGrau Gotés, Mª ÁngelaGrau Sánchez, MiguelMontseny Masip, EduardSobrevilla Frisón, PilarS’apliquen les tècniques clàssiques de propagació de l’error a la transformació de l’espai de color RGB en l’espai de color HSV a un conjunt de 1098 imatges test. El conjunt d’imatges test són 183 paletes de color i sis nivells d’il·luminació diferents. Els resultats que es presenten indiquen com varien la mitjana i la variància per la transformació.