Missions

Preparation of a PhD

Activities

The objective of this thesis is to contribute to the development of new collaborative localization methods, with prior maps, able to manage the scalability of the map with the aim of improving localization accuracy and integrity [1] [2] [3]. In this work, the map will consist of buildings and will be associated with 3D lidar measurements to localize vehicles. The prior map can be obtained from OpenStreetMap (OSM), for example. Collaboration will take place when at least two vehicles observe the same landmarks through direct communication between them.

Collaboration between vehicles is a key element to improve accuracy, fault tolerance and external integrity (real value to be within the confidence interval) of their state estimations thanks to more and redundant information. Collaboration takes place when two vehicles observe the same landmarks and, in this case, an estimation of the relative poses between the vehicles can be carried out indirectly.

To limit as much as possible the information exchanged between vehicles, we aim to represent the buildings map in the form of stable characteristics, which must be robust to changes in perception angles and environmental conditions. The choice of characteristics and descriptors of a lidar point cloud has been the subject of several research studies [4-8]. In this thesis, particular attention will be paid to the choice of characteristics to extract after a detailed state of the art of existing approaches and their limitations. The right choice of characteristics will make it easier to resolve the problem of map scalability. Indeed, the prior maps present imperfections which must be taken into account over time. Thus, if characteristics are detected and not present in the map, then they will be added and conversely the characteristics of the map which no longer exist will be deleted. Thanks to these treatments, autonomy capabilities can be improved in real time.

For the fusion architecture, preference will be given to a distributed architecture. In this case, the vehicles localize themselves in a cooperative way thanks to direct communication between them (decentralized coollaborative localization) to better estimate their poses and improve their own local maps consisting of only part of the environment. A central server is responsible for updating an optimized global map but at reduced frequency. This global update of the map can be carried out through an optimization approach. The approach will then be based on a filtering part (model-based) and another on optimization (for the global map). The approach will then be based on a filtering part (model-based) and another on optimization (for the global map). Indeed, using only a filtering-based approach creates problems when merging maps and can lead to inconsistency problems which affect the integrity of the system. Likewise, using the whole map in the state vector can lead to a heavy computational load.


Regarding the cooperative localization, two approaches can be considered. The first, which we studied in Maxime Escourrou's PhD, consists of comparing the points cloud with the map to generate an observation. This approach requires the choice of a relevant and comparable observation between vehicles. A second approach that can be considered in this PhD consists of a joint comparison of the characteristics of the point clouds of several vehicles. The approach will then be divided into two parts: a part of direct comparison with the map and another between the vehicles.

Regarding the multi-sensor data fusion method, preference will be given to Bayesian estimation methods such as the UKF (Unscented Kalman Filter) and the Schmidt Kalman Filter.

Regarding integrity, each robot can improve its state estimation thanks to shared information and measurements. Particular attention will be paid to the quality of information exchanged between vehicles to improve the quality of estimates. To limit errors as much as possible, a diagnosis step will be added in which thresholds will be set based on learning approaches. As the choice of covariances associated with measurement noise also has a direct influence on integrity, approaches based on learning to automatically adjust the covariances will be considered. This part will be studied in detail, as well as the link between perception uncertainty and localization integrity.


The tests and evaluations will done on the laboratory platform involving three vehicles equipped with 3D lidar sensors.

Skills

Engineering degree and master's degree in robotics and/or computer science and/or automation, good programming skills (Python, C++, matlab). Experimentation and embedded algorithms.

Work Context

Location: Heudiasyc Laboratory, Université de Technologie of Compiègne, France.
The work is part of SIVALab (Renault/Heudiasyc/CNRS) joint laboratory.

The position is located in a sector under the protection of scientific and technical potential (PPST), and therefore requires, in accordance with the regulations, that your arrival is authorized by the competent authority of the MESR.