Safe and accurate localization solution for autonomous vehicles travelling in a constrained environment – road/rail – LOCSP

The first on fault-tolerant hybrid and collaborative solutions for safe autonomous navigation. The coupling of techniques based on both models and data has made it possible to produce a sensor fault-tolerant collaborative localisation solution. Given the stochastic nature of the measurements, we have chosen the informational formalism, which provides measures of dissimilarity between probability distributions called divergences. We use the Jensen-Shannon divergence to synthesise fault indicators, the residuals. The thresholding of these residuals can then be used to detect and isolate sensor faults. We also studied the contribution of learning to diagnostic decision-making. Two models, one for detection and the other for isolation, were trained using various machine learning tools (multilayer perceptron, decision tree and logistic regression).

Cooperation between the vehicles led to the implementation of a decentralised architecture for multi-sensor data fusion and diagnostics. This inter-vehicle cooperative aspect provides informational redundancy, helping to improve the performance of pose estimation and diagnosis. The data from this architecture was used to set up a federated learning paradigm.

The proposed methods have been developed, tested and evaluated on a set of scenarios with real and injected sensor faults.

The second focuses on methods for detecting and reducing electromagnetic interference on GNSS solutions. We began by evaluating the performance of three techniques: Karhunen Loeve Transform (KLT), Adaptive Notch Filter (ANF) and Wavelet Packet Decomposition (WPD), which represent different families of methods, studying in particular the performance indicators associated with safety applications. Based on

the results observed during the state-of-the-art implementation phase, we have proposed

a parameter optimisation methodology for the ANF. This optimisation is performed for linear chirp jammers and shows that an optimal parameter setting has an impact on performance criteria such as accuracy, availability and security. This work contributes to an in-depth understanding of the phenomenon of phenomenon in critical applications. It lays the foundations for the targeted implementation of multi-channel, real-time implementation of an interference detection and mitigation solution. The proposed system aims to continuously detect, classify and characterise different interference to adapt to multiple classes of interferers, sources and power levels.

The implementation of a data acquisition platform has enabled this scientific work to be validated on real data.

As part of the project, a reference database has been compiled. It is representative of the different environments traversed (rural, urban, forest, etc.) and covers the different faults encountered in transport environments: multipath, reflected paths in the absence of a direct path, and interference. These bases were acquired thanks to the equipped vehicle used in the project.

One of the key points when collecting geolocation data, especially when the aim is to use it for post-processing using data fusion algorithms, is the capacity and accuracy of time synchronisation of the data collected. Indeed, given the differences in sampling rates, timestamping, data formats and the diversity of the collection equipment used, ensuring accurate synchronisation is crucial for the post-processing and fusion of data from multiple sources. The project has enabled the methodology and equipment configuration to be developed in order to carry out efficient collection campaigns with effective data synchronisation based on a ROS architecture.

Beyond the synchronisation aspect, a second technical difficulty lies in the need to supplement data collection with a detailed description of the environment and collection conditions. The direct environment has a major impact on the reception conditions and performance of the sensors (particularly the GNSS sensor). Knowledge of the collection conditions is therefore an important factor in ensuring that the data collected can be used effectively.

Once it has been made public (expected in the short term), this database will make it possible to compare the performance of different location solutions on a reference database for the scientific and industrial communities.

With two theses, the project is also proposing new solutions for the detection and reduction of interference and a method for the collaborative localisation of a group of vehicles, which is also capable of detecting and isolating sensor faults, with a view to excluding them from the estimation of vehicle positioning. These results have been the subject of a number of scientific publications.

LOCSP has enabled the teams to develop their skills and tools in the field of localisation. New projects have emerged to extend the scope of the research to new families of estimators and new sources of failure (e.g. spoofing) and to transfer the results to railway applications.

M3 Systems has been involved for several years in the preparation and development of the EN16803 standard, which aims to define a standard for the testing and validation of geolocation systems in the automotive sector. Among the methodologies and tools mentioned in the standard, the use of Record/Playback is an important element.

We are convinced that this approach is just as valid in the rail sector. In this sense, the tools and collection methodology used as part of the project offer very interesting prospects for reuse in order to continue collecting data and increase the volume and diversity of the data available for both the automotive and rail sectors.

Beyond the technical aspect of things, there is also a commercial interest for M3 Systems in offering its customers the possibility of having sensor data that can be directly used in replay by STELLA test tools.

Adaptative Diagnosis for Fault Tolerant Data Fusion Based on alpha-Rényi Divergence Strategy for Vehicle Localization, M. Khoder, N. Ait Tmazirte, M. El Badaoui El Najjar, N. Moubayed. MDPI - Entropy, April 2021. DOI: 10.3390/e23040463

Navigation Context Adaptive Fault Detection and Exclusion Strategy based On Deep Learning & Information Theory: Application To a GNSS/IMU integration, Nesrine Harbaoui, Nourdine Ait Tmazirte, Khoder Makkawi, Maan El Badaoui El Najjar, ION GNSS+, Sept. 2021

Revolution of the transportation systems (evolving toward autonomous cars or trains in a world where digitalization will be everywhere) implies the development of a panel of embedded technologies among which localization and communication are key elements. In this framework, LOCSP (Solution de LOCalisation Sûre et Précise) will focus on the localization function and in particular on the requirement for accurate and safe positions. Indeed, the new positioning solutions shall provide on one hand the minimum accuracy required by the application (road level, track identification, position of the vehicle on a track…) and, on the other hand, ensure the user that the position can be trusted through the development of integrity monitoring solutions. Most of the solutions developed exploit the benefits of heterogeneous sensor fusion for a better robustness and accuracy.
LOCSP gathers two research teams (IFSTTAR and CRIStAL) and one SME (M3 Systems) in order to develop and compare safe and accurate localization solutions even in presence of faults and thus, answer the two following challenges : make the position more robust and accurate by implementing the most recent technologies and hybridizations techniques ; Allow everyone to benefit from a complete and detailed evaluation of the different solutions.
LOCSP intends, on the one hand, to complement existing studies for better performance, and, on the other hand, answer to need of comparison platforms of any algorithms developed in a different team.
Researches leaded in the framework of LOCSP will bring contributions on the two following topics in particular :
- First, on fault tolerant data fusion and cooperative solutions. It will consist in developing a framework for Fault Detection and Exclusion (FDE) for a Collaborative Localization (CL) of a multi-vehicle fleet, equipped with multi-sensors solutions.
- Second LOCSP will investigate the different techniques for interference detection and mitigation, evaluate and compare their respective performance. In particular, we will investigate the capacity of these different techniques to face RFI and in particular, the one from RAIM algorithms (Receiver Autonomous Integrity Monitoring) as well as the behavior of fusion algorithms in a PhD work. M3Systems will provide the tools for interference injection in the dataset.
In the project proposed, a reference database will be composed, as a representative set of crossed environments (rural, urban, forests) and feared events (multipath, reflected paths, interferences… These data will result from previous campaigns collected by the consortium members and will be complemented by specific other campaigns. Once public, this database shall allow everyone (scientific and industrial communities) to compare performance of positioning solutions on a common frame.
LOCSP objectives are then: (i) to compare performance reached by existing solutions; (ii) to propose a panel of solutions and algorithms with growing costs and complexity, in order to analyzed their behavior in the presence of faults in order to define their limits and strengths.