MSCA-COFUND-CLEAR-Doc-PhD Position#CD22-61: Towards a Reliable and Resilient Positioning System for Soft Mobility with a Green AI Approach

Mobility and transport play crucial roles in achieving sustainable development goals. With the recently emerged new form of urban soft mobility (from pedestrians, e-scooters, e-bikes to the self-balancing human transporter vehicle), more and more users choose these new modes of mobility as alternatives to replace single-occupied vehicles or for last-mile solutions. Soft mobility can no doubt contribute to the sustainability of society: less congestion, more environmentally friendly and less noise in the city. However, the users of soft mobility are vulnerable compared to other transport users, where accidents can be easily fatal once happen. Based on the World Health Organization’s report on Road Safety (2021), around half a million pedestrians, cyclists or users of motorized two-wheelers collectively known as vulnerable road users, die yearly in road traffic crashes. The objective of this PhD thesis is to enhance the safety of soft mobility users by providing a more reliable and resilient positioning system assisted by AI techniques.

The first challenge of this PhD thesis is to better understand the risk and criticality of the soft mobility users, or more generally, the vulnerable road users. This would include the definition of the critical use cases for soft mobility, which leads to consolidated positioning user requirements (accuracy, availability, integrity, continuity, Time-to-Alert, maximal tolerable positioning error - Alert Limit). These requirements will be used to warn users in the case of no reliable information provided by the positioning systems. Then the second challenge is to qualify sensor measurements (such as GNSS, INS and Magnetometer), which aims at detecting the faulty measurement and then excluding or repairing them. The traditional statistical-based Fault Detection and Exclusion (FDE) is based on error modeling and consistency checking through measurement redundancy [1]. The main limitations of these traditional methods are first, the complexity of modeling measurement errors in harsh environments (such as urban canyons for GNSS); second, in the case where the majority of the measurements are erroneous, the consistency checking will fail if no other special measures are taken. That is the reason why a data-driven approach will be used to handle this problem in this PhD thesis. By using in priority the existing labeled data collected previously by the GEOLOC Laboratory as well as the transfer learning technique based on the pre-trained model, a Green AI approach will be followed by reducing the computational cost and the carbon emission. For example, the transfer learning technique can be applied to train the model from the vehicular data and then transfer it to soft mobility applications. Once the sensor measurement is qualified, the uncertainty of the positioning solution will be finally estimated to bound the position error with the integrity risk pre-defined in the first challenge for different vulnerable mobility user profiles. The AI regression methods will be used to better solve the problem of conservative positioning error bounds calculated by traditional statistical methods. Also, the application of more developed machine learning methods for solving the problem will be addressed.

In this way, an end-to-end reliable and resilient positioning system will be designed, which can provide timely warning to users in the case of misleading positioning information provided by the system. This functioning is essential to improve the safety of vulnerable road users. By enhancing the safety and accessibility of the soft mobility positioning systems, more and more people will be encouraged to take these efficient and green mobility modes. This will help to achieve the sustainable development goals (SDGs) and build a better future for the planet.

This PhD will be co-supervised by researchers from the GEOLOC Laboratory of the University Gustave Eiffel in France and the Department of Computer Science of the University of Helsinki. GEOLOC laboratory has the expertise in positioning with multisensor hybridization for multimodal transportation (vehicles, pedestrians and other soft mobility) assisted by AI [2]. And the team of the University of Helsinki has expertise in AI-based positioning algorithms, mainly deep learning, combined with computer vision techniques [3-5]. The PhD student will spend 6 months as visiting researcher at the University of Helsinki.

Required capacity: signal processing, artificial intelligence, multisensory fusion positioning, state estimation, python, matlab

[1] Zhu, Ni, et al. ""Evaluation and comparison of GNSS navigation algorithms including FDE for urban transport applications."" Proceedings of the 2017 International Technical Meeting (ITM) of The Institute of Navigation (ION). 2017.

[2] Kone Y, Zhu N, Renaudin V, et al. Machine learning-based zero-velocity detection for inertial pedestrian navigation[J]. IEEE Sensors Journal, 2020, 20(20): 12343-12353.

[3] Kia, Ghazaleh, Laura Ruotsalainen, and Jukka Talvitie. ""A CNN Approach for 5G mmWave Positioning Using Beamformed CSI Measurements."" Proceedings of the International Conference on Localization and GNSS (ICL GNSS) 2022.

[4] Al-Tahmeesschi, A., Talvitie, J., López–Benítez, M., & Ruotsalainen, L. (2022, June). Deep Learning-based Fingerprinting for Outdoor UE Positioning Utilising Spatially Correlated RSSs of 5G Networks. In 2022 International Conference on Localization and GNSS (ICL-GNSS) (pp. 1-7). IEEE.

[5] L Ruotsalainen, A Morrison, M Mäkelä, J Rantanen, N Sokolova (2021). Improving Computer Vision Based Perception for Collaborative Indoor Navigation. IEEE Sensors Journal.