Guillaume Jornod – A visualization tool for Particle Swarm Optimization and other machine learning techniques
Supervisors: Prof. Alcherio Martinoli, Dr. Inaki Navarro and Ezequiel Di Mario,
About The Distributed Intelligent Systems and Algorithms Laboratory (DISAL), EPF, Lausanne
We are interested in machine learning optimization techniques for distributed learning applied to mobile robotics. In particular, we intend to focus on Particle Swarm Optimization, a relatively new metaheuristic that has shown promising results when compared to more traditional techniques such as Genetic Algorithms. Optimization techniques have a set of parameters that are critical to the performance of the algorithm, and can be related to hardware constraints when applied to distributed robotic platforms, such as neighborhood topology and communication range. The goal of this project is to develop a visualization tool that displays the evolution of the set of candidate solutions in canonical benchmarks with and without added noise, with the purpose of gaining insight on the effect of these parameters on the performance of the swarm.
Emmanuel Lanti – Implicit time advance of shear Alfven waves in magnetic plasmas
Supervisor: Prof. Paolo Ricci
About Centre de Recherches en Physique des Plasmas (CRPP), EPF, Lausanne
A global fluid code named GBS is being developped at the CRPP to study turbulent transport in the open magnetic field line region at the very edge of magnetically confined fusion plasmas such as Tokamaks. Due to the strong collisionality in this region, one uses Bragsinskii fluid equations, which evolve the density, electron temperature, electrostatic potential, ion and electron parallel (to the magneti field) velocities. These fields are subject to Coulomb collisions and electromagnetic fields. The code evolves these fluid quantities on a 3D grid using centered finite differences an an explicit fourth order Runge Kutta integration scheme.
The code has currently two main drawbacks. First, plasma turbulence is strongly anisotropic and turbulence tends to have long wavelengths parallel to the magnetic field and short ones in the perpendicular direction. Consequently, the accuracy of the parallel gradient operator must be carefully monitored. Second, Braginskii equations contain the electrostatic shear-Alfven wave, which propagate field at frequencies much higher than the turbulent frequencies of interest. Therefore, the time step is constrained by a relatively severe CFL condition.
One way to overcome these limitations is to use an implicit time advance for the parallel advance. On one hand, implicit schemes are supposedly more costly than explicit ones but on the other hand they should allow to increase the time step significantly. The goal of the project is therefore to implement a variety of (semi-)implicit schemes in a simplified model that evolves the linear propagation of shear-Alfven waves subject to parallel diffusion.
Léonard von Niederhäusern – Optimizing the gains of a hydro power station: different approaches
Supervisor: Prof. Friedrich Eisenbrand
Chair of Discrete Optimization – DISOPT, EPF, Lausanne
A hydro power station basically consists of two water basins relied by a pump/turbine which is able to either bring water to the upper basin or to produce electricity as the water flows down to the lower basin. Depending on the energy price, we are looking for the best pumping/turbining profile in order to maximize the gains of an energy company. Different approaches are going to be explored. The creation of a computer program that would be able to compute such an optimal profile is on an a priori to-do list.
Adrien Gaudard – Multi-Robot Relative Localization Using 2D Laser Range Finders
Supervisors: Dr. Andreas Breitenmoser, José Nuno Pereira
The Distributed Intelligent Systems and Algorithms Laboratory (DISAL), EPF, Lausanne
The MOnarCH project  studies the use of networked robots in a children’s hospital for supporting the medical staff and for interacting with the children and their visitors. As a part of their tasks, the robots need to cooperatively patrol the corridors and wards of the hospital from time to time in order to monitor the situation or to look for children.
The goal of this student project is to implement a cooperative localization strategy based on the robots’ onboard 2D laser range finders , aiming at the relative localization (range and bearing) of multiple robots in the presence of sensing uncertainty. The student will work with ROS  and the physics-based robot simulator Webots . The relative localization will include the detection of the other robots from the recorded laser point clouds and the distributed estimation of the robots’ relative poses; the estimation may be based on the fusion of the robots’ wheel odometry and the laser range finder measurements. The localization strategy will be tested and demonstrated with a group of simulated robots under different simulation settings.
Christian Kanesan – HydroContest: Hydrofoil Geometry Optimisation Using Numerical Flow Simulation
Supervisors: Dr. Dede’ Luca and Dr. Simone Deparis
Chair of Modelling and Scientific Computing – CMCS, EPF, Lausanne
Andrea Di Blasio – Numerical homogenization and inverse problem
Supervisor: Prof. Assyr Abdulle
Elena Queirolo – Low rank preconditioners
Supervisor: Prof. Jan S. Hesthaven
Joël Répond – Preheating with non-canonical kinetic terms
Supervisor: Prof. Mikhail Shapochnikov
Joël Repond CSE project will be devoted to the numerical study of the preheating stage1
in models with non-minimal kinetic terms in a 3+1 expanding Universe. This kind of terms
appears naturally in string theory models with non-trivial geometries for the K¨ahler potential
 or in scale-invariant extensions of the Standard Model as those considered for instance in
Refs. [2, 3]. The effect of non-canonical kinetic terms in the particle production after inflation
has been extensively studied in the literature [4, 5, 6, 7]. These analytical studies are however
quite limited and apply only to the beginning of the preheating stage, where the backreaction
effects can be safely neglected. The evolution of the perturbations after backreaction is
very non-linear and non-perturbative and numerical simulations on the spacial lattice are
eventually needed. There exist several numerical codes allowing to evolve interacting scalar
fields in an expanding universe [8, 9, 10]. However, these codes are based in symplectic
integration routines and are therefore not well-suited for the simulation of non-canonical
kinetic terms, which involve the product of the field and its time derivative. The existing
codes must be therefore modified in order to store both the field and the field derivative at
the same time during the time step. The alternative integration routine is expected to be
slower and to require more physical memory than the usual symplectic integrators and a
further parallelization of the program might be eventually needed.
 C. P. Burgess, Class. Quant. Grav. 24 (2007) S795 [PoS P 2GC (2006) 008] [PoS
CARGESE 2007 (2007) 003] [arXiv:0708.2865 [hep-th]].
1- Preheating is the period after inflation during which the energy stored in the inflaton is transferred to
the matter and radiation of which our universe is made up.
 M. Shaposhnikov and D. Zenhausern, Phys. Lett. B 671 (2009) 187 [arXiv:0809.3395
 J. Garcia-Bellido, J. Rubio, M. Shaposhnikov and D. Zenhausern,
Higgs-Dilaton Cosmology: From the Early to the Late Universe
Phys. Rev. D 84 (2011) 123504 [arXiv:1107.2163 [hep-ph]].
 J. Lachapelle and R. H. Brandenberger, JCAP 0904 (2009) 020 [arXiv:0808.0936 [hepth]].
 T. Matsuda, JHEP 0810 (2008) 089 [arXiv:0810.3291 [hep-ph]].
 N. Bouatta, A. -C. Davis, R. H. Ribeiro and D. Seery, JCAP 1009 (2010) 011
 J. Garcia-Bellido, J. Rubio and M. Shaposhnikov
Higgs-Dilaton cosmology: Are there extra relativistic species?
Phys. Lett. B 718 (2012) pp. 507-511 [arXiv:1209.2119 [hep-ph]],
 G. N. Felder and I. Tkachev, Comput. Phys. Commun. 178 (2008) 929 [hepph/
 A. V. Frolov, JCAP 0811 (2008) 009 [arXiv:0809.4904 [hep-ph]].
 R. Easther, H. Finkel and N. Roth, JCAP 1010 (2010) 025 [arXiv:1005.1921 [astroph.
Hugo Babel – Mobility patterns at Lausanne train station
Supervisor: Prof. Michel Bierlaire
Pedestrians, contrary to the other modes, do not have defined network and their motion is not constrained by strict rules. A characteristic feature of pedestrian route choice is that routes are continuous trajectories in time and space, and pedestrians choose a route from an infinite set of alternatives. As a consequence we have observed the distribution of walked distances and travel times for specific origin-destination pairs at the train station in Laussane.
The goal of this project is to develop a model describing the deviation from the shortest-path by performing further analysis of spatial and temporal motion patterns of pedestrians in mobility hubs. For this purpose the great potential of the available unique data set will be exploited. The data set contains pedestrian trajectories collected at the train station in Lausanne using a large- scale network of smart sensors that automatically locate pedestrians in space and to track them across time. The temporal resolution leads to thousands of points per pedestrian enabling the detailed study of the different types of pedestrian flows in public spaces.
Federico Martinez Lopez – 3D surface reconstruction from photometric stereo
Supervisors: Dr Loïc Baboulaz and Ph. D. Niranjan Thanikachalam (LCAV)
3D surface reconstruction is a fundamental problem in Computer Vision. Common approaches to tackle down the problem are: conventional stereo, shape from shading, structured light and photometric stereo among others. Photometric stereo is an image-based method used to recover the shape of an object from a set of images taken under various illuminations and a fixed viewpoint. Under the lambertian reflectance model and suitable assumptions of the image formation process we can construct the reflectance maps that give information about the geometry of the objects in the scene. In this project we investigate two methods for the estimation of the surface normals: the direct inversion of the lambertian model and the recovery of the underlying low-rank matrix to discard noise (shadows, specularities) in the acquired images. Next, for the reconstruction of the 3D surface we compare the results of a local and a global integration methods.
This project is part of the eFacsimile project that focuses on high-quality Artwork acquisition and reproduction for modern digital devices such as laptops, smartphones and tablets. By faithfully capturing and representing digitally all the various features that compose a visual artwork (e.g. colors, textures, light reflectance), the viewing experience can be greatly enhanced. Combined with an ergonomic interface, the user can then interact and actively manipulate the digital replica in a completely new way.
Christos Kotsalos- Bath Tub Vortex : Stability & Sensitivity Analysis
Supervisor: Prof. François Gallaire
The bath-tub vortex is one of the most familiar fluid dynamical phenomena. It is the intense and concentrated swirling motion around a hole located at the bottom of a symmetric tank. After detailed investigations, the physical origin of its appearance is still unclear. There are two candidates as the source of this phenomenon. The first is an imperceptible but non-zero rotation in the initial condition and the second is that the tank through the viscous boundary layer provides the torque necessary to set in the swirling motion. The Coriolis effect seems to play not an important role to the appearance of the vortex. The goal of this project is to investigate the above candidates and to understand the mechanism of the vortex creation. In the first part, we compute the base flow which is the solution of the steady and axisymmetric Navier-Stokes equations without swirl. In the second part, we do a stability analysis of the base flow. Essentially, we introduce a perturbation in the steady field and investigate whether it grows or decays. From this step, we examine if the base flow is naturally unstable and consequently the vortex is created. The stability analysis results to a stable base flow. For that reason, in the third part we do a sensitivity analysis so as to examine how sensitive is the flow domain to modifications of the base flow. Finally, in the last part we investigate some case studies where we change slightly the geometry and destabilize the flow field by perturbing the base flow. We conclude that the vortex cannot be created in a perfect flow field (symmetric tank and flow conditions) and that the asymmetries are the key factor for the presence of the Bath-Tub Vortex.
Benjamin Paccaud – Development of a material class inside the C++ Finite Elements library Akantu
Supervisors : Pr. Jean-François Molinari and Dr. Ramin Aghababaei
Akantu is a C++ library designed for finite elements and developed at the Computational Solid Mechanics Laboratory at EPFL. It is mainly designed for solid mechanics but has developments in heat transfer and contact for example. In this project, a first step was to get familiar with the library and some existing material classes by comparing the numerical results and the analytical solutions. The second step was to write a new material class for a large-deformation plasticity model with isotropic hardening. The results were then compared with a commercial software for two typical problems of solid mechanics : a uniform traction test and a plate with a hole.
Miryam Chaabouni – Mars Image Registration to track changes on Mars
Supervisors : Dir. Volker Gass and Dr Anton Ivanov
Swiss Space Center, EPFL
The surface of Mars shows particular circular pits often called ‘Swiss-cheese features’ which can only be found on the south pole. The Mars Reconnaissance Orbiter provides very large, high resolution images (HiRISE) of huge surfaces that need to be processed computationnally. The goal of the project is to find the most efficient way to peform matching of two different imgaes of the same area, taken from different angles and with different illumination conditions. Additionnaly, the specific circular shapes should be detected and matched. A very robust feature detector called SIFT, was chosen for the project; and its GPU version SiftGPU was used to perform the tasks required for the project. A C++ code based on SiftGPU was written in Visual Studio, and tested on HiRISE images.
Lucas Daniel Amoudruz – Refining leg-movement extraction in high-speed videos of insect locomotion
Supervisor : Prof. Unser
Image segmentation is widely used in Biology. Active contours (commonly called snakes) are efficient tools in this domain, and work well if they are close. Nevertheless, open snakes are more adapted for particular shapes, as fly legs, but often make loops or have undesired behaviors. The goal of this project was to design efficient methods to avoid these behaviors, as new snake energies or new conditions.
It can be improved by first computing the distance transform of the input image. This method is efficient in most cases, but stillallows loops in particular case. One can then develop a condition to avoid loops, which gives good results.Finally, for sequences of images, we used a ‘movement’ energy which tends to keep a good parameterization of the snake.
Joseph Lemaitre – Dynamic Platooning for Intelligent Vehicles
Supervisor : Prof. Alcherio Martinoli
Grouping smart cars into platoons has become important for the automotive industry due to the advantages it brings in, such as reduction of fuel consumption, increasing the safety and comfort of traveling, and increasing capacity of roads. This project re-implement a work of Gowal et al. on leaderless graph-based platooning and brings the following two improvement : First, the vehicles are able to join or leave the platoon without jeopardizing the stability of the formation. This is necessary to enjoy the benefits of platooning. Vehicles have individual origins and destination, so traveling in a formation is only a part of their trip. All the cars have the same simple local behavior that maintain the globally the shape of the platoon. A communication layer is made available for joining/leaving requests. Second, for use in a realistic scenario, we want a formation that can follow any road, straight or curved. The authors of the graph-based algorithm assumed that the x-axis of the car’s coordinate system of the car is always aligned with the road direction. Therefore, they either needed a straight road or the a-priori knowledge of the path so they could use curvilinear coordinates. To overcome these requirements, we added a lane keeping behavior (through a computer vision line-following algorithm) on top of the graph-based platooning. Those modifications were implemented and tested using the Webots robotics simulator.
Ryan Giroud – Flow around a hydrofoil in harmonic vibration. CFD simulations and validation of a Vortex-Lattice and Doublet-Lattice code
Supervisor : Dr Joël Cugnoni
For the Hydoptère project, the LMAF has developed a Matlab suite of potential solvers based on Vortex Lattice, 3D Panel and Doublet Lattice to perform quick simulations of hydrodynamic added mass effects as well as lift calculation in static flow and harmonic vibration. In principle, these solvers allow to quickly evaluate the performance of hydrofoils in the case of an attached flow et have been integrated into a routine for computing the dynamic stability of coupled fluid-structure problems using the PK method.
The purpose of this project is to develop a Navier-Stokes CFD model of the flow around a reference hydrofoil (NACA0009) for steady flow conditions and under harmonic disturbances of the hydrofoil according to one of its eigen mode in air. The results of these simulations in term of lift and moments as well as modal forces are compared with the predictions made using the potential codes previously developed to estimate their reliability and validate their predictions.
CFD simulations are performed on the open source finite volume code Code_Saturne in ALE formulation. The vibration of the foil is represented by an ALE velocity field on the boundary according to the studied mode of the structure and implemented into a Fortran user-subroutine. The flow conditions are chosen to match the tests performed prior at LMH and only small perturbations / angles of attack are considered in order to validate the potential code in its expected validity domain. The mesh is optimized to minimize the computation time. The main objective is to achieve convergence on the lift at low angle of attack in attached flow conditions.