Skip to main content

Job Offer


PhD: Learning and recognizing virtual beacons to improve the mobility of visually impaired persons

Do you have a heart for the contribution of artificial intelligence in the construction of physical models, to address the challenges and needs of new mobility services? The inmob joint laboratory on the Nantes campus of the Gustave Eiffel University of France could be the next step in your career.


Research at the Geoloc Laboratory of the Gustave Eiffel University focuses on the development of dynamic positioning methods and systems for travellers. It owns a wide range of skills in the design and programming of beaconless positioning and navigation algorithms with original testing facilities. The Okeenea Digital company is an innovative company from the Okeenea group, specialised for 26 years in the design and supply of accessibility solutions to improve mobility in situations of disability. These two organisations have joined forces in the joint laboratory inmob (disability mapping with INertial signal to facilitate MOBility) to improve the mobility of people with disabilities.

Our ambition is to design new approaches based on artificial intelligence to build physical models that describe the individual specificities of movement of people with disabilities and thus individualise the positioning approaches. This research will increase the performance of current geolocation technologies, which are still insufficient to assist a blind or hearing-impaired person throughout his or her route in complete safety without depending on a very dense network of beacons. To strengthen the labcom team and contribute to the work in artificial intelligence for improved mobility services, we are looking for a research engineer or postdoc, ideally with complementary skills in physics/mathematics and artificial intelligence.

Research topic

Many smartphone-based applications help people navigate. They use GPS, Wi-Fi, Bluetooth, mobile phone data, camera and inertial signals to compute the mobility path consisting of geographic coordinates. But their accuracy is still insufficient for applications where accidents are possible. The estimated routes suffer from drift and positioning errors due to the low quality of the sensors embedded in the smartphone and surrounding obstacles. Beacons (Bluetooth tags, other Internet of Things tags) can be used to improve the performance, but they require a specific infrastructure of beacons deployed in the city. It is also possible to consider virtual points of interest, such as transitions between outside and inside or up/down stairs, to correct the coordinates using map data and the content of the associated geographic information system. Although this concept is well known, it lacks robustness and is not adapted to the specific needs of people with disabilities. Like the applications that integrate the use-case needs (e.g. Soundscape) at the heart of the R&D, this thesis aims to observe the behaviour of visually impaired people to propose new virtual beacons that are specific to them.

Research objectives

This thesis aims at learning the mobility patterns of visually impaired people to build databases of virtual beacons for opportune recalibration of the estimated foot-track in real-time on a smartphone. The work will begin by reading and analysing articles on methods for recognising displacement patterns from signals and images. Existing map matching and pattern recognition codes for pedestrian dead-reckoning algorithms will be redesigned to identify the need for virtual beacon registration. Observing and identifying through experiments the virtual points of interest related to the situation of the disabled and more particularly the visually impaired will be a key stage of the research. It will then be necessary to propose and test automatic learning or deep learning methods to recognise these points of interest. The final objective is to bring these approaches to smartphones to improve and secure the guidance of visually impaired people in multimodal transport.

Required Skills

  • MSc/Eng Degree in signal processing, geomatics, artificial intelligence or computer science
  • Knowledge of signal processing, data mining, applied mathematics, machine learning methods and statistics
  • Knowledge of multi-sensor fusion positioning and estimation theory
  • Strong programming skills in Python, Matlab and TensorFlow (cpp preferred)
  • Experience of working with machine learning applied to physics
  • Proven ability to integrate large-scale data
  • Understanding of connected object technologies and Android programming
  • Ability to effectively integrate research results into publications and intellectual property
  • Creativity and openness to innovation
  • Enthusiasm, responsibility and excellent collaboration skills
  • Passion for producing high quality data
  • Strong oral and written skills in English
Contract TypePhD, 3 years full time (38h30/week) with a start in October 2021 (flexible date)
LocationGEOLOC Laboratory, University Gustave Eiffel, Nantes, France
Deadline31/12/2021 23:50 - Europe/Brussels
ApplicationSend all documents (cover letter, CV, diplomas, list of productions / publications, references) in a single pdf document to


Intention Analysis of Vulnerable Road Users


In recent years, a great research interest came up on the automatic protection of vulnerable road users (VRUs) like pedestrians, bicyclists, or even cars. One main problem of it is the estimation of the position and additionally the intention of all road users in order to avoid upcoming collisions. Positioning in urban canyons and underground environments like tunnels, where GNSS reception is difficult, will require a combination of sensors and other information such as environmental maps in order to overcome positioning service failures. A solution for pedestrian navigation developed at Univ-Eiffel and DLR is pedestrian dead reckoning based on inertial measurements with a body worn sensor. Unfortunately, those systems relying only on the IMU still suffer from a remaining drift.  Therefore, the developed pedestrian navigation systems reduce the remaining drift by different techniques like learning a map of the environment, feature detection or GNSS, when available. With these techniques accurate tracks of pedestrians can be estimated. But still the intention of the vulnerable road user - i.e. where the road users will most probably go next and the next destination of the user – is not yet fully investigated. The overall goal is to locate all vulnerable road users, estimate their intention, and inform the nearby road traffic for instance at intersections in order to create awareness of approaching VRUs and to enable collision avoidance.


There are different ways of estimating the intention of the vulnerable road user: One way is to make use of the last estimated track and to predict the future track from it. For this, machine learning techniques can be for instance used in order to learn the possible continuing of the tracks of the road user depending on the track history. Another way is to combine the estimated tracks of multiple pedestrians, bicyclists, and cars in a suitable way. From the combination of all tracks a layout of the environment can be learned that enables the generation of a map containing not only the road map but also the environmental details of the area. In addition, this map can provide information about obstacles like potholes or road construction and determine possible destinations or intentions of the road users.


Your task is to predict reliably the next steps/path of the road user out of the knowledge of obstacles on the road and the possible destination of the user in order to provide information about possible collisions in the environment. More specifically, your tasks are:

  1. Do a literature research on track prediction and intention analyzing methods.
  2. Develop a machine learning technique for track prediction based on the trend of the previous track. The machine learning technique shall be designed for pedestrian tracks, which include all possible turns and random movements.
  3. Analyze the pedestrian positioning performance requirements (e.g., accuracy, maximal tolerable positioning errors etc.) in different environments and try to apply these criteria as constraints for track/intention prediction.
  4. Compare the developed technique to methods given in the literature.
  5. Combine multiple tracks in order to obtain an environmental map. Develop a method to use that map for track prediction and intention analysis.
  6. Compare and evaluate the proposed methods with real experiments by using the positioning equipment existing in the labs.
  7. Opportunity to publish in conferences or journals.


  • Knowledge of signal processing, machine learning and statistical analysis;
  • Strong programming skills in Matlab, Python;
  • Creativity, responsibility and strong oral/written skills in English.
Contract TypeInternship, Gross salary 1100 / month
LocationWorking on the site of Munich in the Vehicular Applications Group, Communications Systems Department, Institute of Communications and Navigation, DLR
Deadline31/01/2022 23:50 - Europe/Brussels
ApplicationPlease send your CV and note transcript to :, and

Evaluation of the Body Mask Effects on GNSS Receiver for Pedestrian Navigation


With the continuous development of intelligent objects, new mobility applications, from pedestrian navigation aids to autonomous vehicles, require highly accurate positioning information. Especially for the reliability-critical applications, such as guidance for blind people, providing accurate positioning information is essential. Global Navigation Satellite Systems (GNSS) are still the main positioning tools for outdoor localization services. However, it is commonly accepted that urban environments present great challenges to common commercial GNSS receivers. This is because GNSS positioning performances can be severely degraded by obstacles, which lead to multipath and Non-Line-of-Sight (NLOS) reception. It is essential to distinguish the LOS/NLOS receptions in order to process the latter in an appropriate manner and thus improve the GNSS positioning accuracy.


In the current literature, much research work has addressed the problematic of NLOS mitigation for vehicular navigation applications. One of the most efficient techniques is to use a sky-pointing fisheye camera to differentiate between the sky and obstacles. Then LOS/NLOS can be classified by projecting the visible satellites onto the previous obtained skymask image [1]. However, when the GNSS receiver is positioned on the body of a pedestrian (hand-held or foot-mounted), not only the environmental obstacles (such as buildings, vegetation, etc) but also the pedestrian’s body can create mask effects, which can potentially degrade the GNSS positioning performances. Very few research works in the state-of-the-art has addressed the body mask effects on GNSS signal reception for pedestrian navigation. The main objective of this internship is to analyze the mask effects of both environments and the pedestrian’s body on a hand-held and foot mounted positioning equipment, which aims at improving the GNSS positioning accuracy.  


This internship consists essentially of simulation work including image processing and integration of multiple data (ephemeris, topographic and altimetric datasets, etc.). Nevertheless, it is planned to carry out - at the beginning of the period - a set of experimental fieldwork aiming at collecting data. In this objective, it will be a question of installing a camera on the top of the target pedestrian positioning equipment named ULISS (Ubiquitous Localization with Inertial Sensors and Satellites), which is developed by the GEOLOC Lab. From the collected datasets, the main tasks of this internship are as follows:

  1. Make image segmentation to differentiate the sky and obstacles;
  2. Choose a proper projective transformation method to project the visible satellites from the ephemeris onto the image;
  3. Based on the previous two steps, analyze the GNSS LOS/NLOS receptions for the hand-held and the foot-mounted positioning equipment. Compare both results to the reference trajectory and draw conclusions regarding the body mask effects.
  4. Develop an accurate 3D urban model of the area, explore the 3D city model, realize a skymask matching technique [2] to make accurate positioning in urban environments (two different techniques for producing the sky maps associated with each position will be explored whether they are raster or vector).
  5. Propose a method to better take advantage of the LOS/NLOS classification in the navigation filter (such as NLOS exclusion or weighting according to LOS/NLOS, etc) and then validate the proposed method by real experimentation.
  6. The work may result in a high-quality scientific paper presented at a conference or submitted to a journal.


[1] Horide, K., Yoshida, A., Hirata, R., Kubo, Y., & Koya, Y. (2019, July). NLOS Satellite Detection Using Fish-Eye Camera and Semantic Segmentation for Improving GNSS Positioning Accuracy in Urban Area. In Proceedings of the ISCIE International Symposium on Stochastic Systems Theory and its Applications (Vol. 2019, pp. 212-217).

[2] Lee, M. J. L., Lee, S., Ng, H. F., & Hsu, L. T. (2020). Skymask matching aided positioning using sky-pointing fisheye camera and 3d city models in urban canyons. Sensors, 20(17), 4728.

Contract TypeInternship monthly salary about 550 €, Flexible beginning dates: from February to April 2022, during 5-6 months
LocationWorking on the campus Nantes of the Univ. Gustave Eiffel and co-supervised by the team Geoloc and the team CRENAU of AAU Lab
Deadline31/01/2022 23:50 - Europe/Brussels
ApplicationPlease send your CV and note transcript to:,,

Positioning uncertainty estimation using machine learning


From autonomous vehicles to connected wearable devices, accurate positioning information is required for all the Location-Based Services (LBS). For the safety-critical applications, however, the reliability of the positioning information is more important. If one can monitor the reliability of the information provided by the positioning system, in case of non-reliable positioning delivered, an alert can be sent to warn the users of the potential dangers. This is the concept of positioning integrity monitoring, which can guarantee the safety of the LBS users. As a hot issue originally introduced in the aviation domain for the Safety-of-Life (SoF) applications, positioning integrity monitoring has been being discussed and designed in recent years for terrestrial applications such as autonomous vehicles. One important technique proposed by the integrity monitoring technique is to estimate a statistical positioning error bound (i.e., Protection Level). This estimated position error can be used as an indicator of system reliability while contributing to the decision-making to prevent users from huge system errors. Therefore, it is essential to estimate the positioning error bound, which can not only properly bound the true position error but also with reasonable sizes, which means it should not be too big.


There are different algorithms in the current literature to estimate the position error bound. Most of them are based on the statistical analysis of source errors’ propagation through the corresponding positioning filter. These methods are generally computational costly and have the risk of providing too huge error bounds, which can impact the continuity performance of the system. Also, some hidden errors or slowly growing errorq can be possibly ignored when calculating the error bounds with the traditional methods. The AI-based positioning uncertainty estimation has recently emerged in some Radio Access Technology (RAT)-based positioning system, which proves to be highly promising. The main objective of this internship is to explore the feasibility of using AI-based methods to estimate the positioning error bound for a hybrid GNSS/INS positioning system.


The main tasks are as follows:

  1. Understand the theories of positioning uncertainty estimation with classic statistical methods;
  2. Do a literature research on AI-based positioning uncertainty estimation and identify the best methodology;
  3. Construct the training database: by using the existing ones or by making new acquisitions;
  4. Train the prediction model and test it;
  5. Compare the proposed AI-based method with the traditional methods.


  • Knowledge of signal processing, machine learning and statistical analysis;
  • Strong programming skills in Matlab, Python;
  • Creativity, responsibility and strong oral/written skills in English.


The work may result in a high quality scientific paper presented at a conference or submitted to a journal.
Opportunity to continue a PhD with the funding of an europeen project.

Contract TypeInternship monthly salary about 550 €, Total duration 5-6 months, Reimbursement of up to 50% of the public transport pass, A very low cost restaurant sponsored by the university
LocationWorking on the campus Nantes of the Univ. Gustave Eiffel
Deadline31/01/2022 23:50 - Europe/Brussels
ApplicationPlease send your CV and note transcript to: