ABSTRACT

For autonomous driving, a false estimation of localization could create hazardous situations and threaten lives. Therefore, it is necessary to increase the reliability of localization systems by enhancing their accuracy and defining and detecting their operating limits. This thesis tackles the integration of Simultaneous Localization And Mapping (SLAM) in an autonomous vehicle for outdoor urban and peri-urban environments under real-life conditions.
SLAM offers immediate localization capabilities while enabling simultaneous construction of a map of the surroundings. SLAM does not require any prior knowledge of the environment and is independent of the infrastructures.
The topics this manuscript addresses mainly emerge from the long-term operation, the diversity, the complexity, the dynamicity and the large-scale of the outdoor environments.
The main topics are: 1. robustness of a SLAM solution by detecting its operating limits. 2. accuracy by alleviating the impact of models’ approximations. 3. scalability and resources awareness using a solid map management technique.
Confusing structures in the environment cause SLAM to fail by misleading its estimation process. SLAM failure is a significant issue that should be taken into account in order to build a robust localization system for autonomous driving. Two approaches to detecting situations in which SLAM may fail are proposed.
The first approach constitutes a relevant descriptor vector analyzing solely raw laser data. Hence, it detects a priori potential failure scenarios which makes it independent of the underlying SLAM implementation.
The second approach exploits the likelihood scores distribution, which makes it rely on the estimation process but independent of the sensor used. This approach operates in parallel to SLAM. The decision in both approaches is made using different Machine Learning models. Approximations in SLAM models (e.g. map representation model, displacement model) induce systematic errors in their estimations. To attenuate such errors; our approach uses two types of relevant information: the previous relative pose estimations, and the likelihood scores distribution. The prediction is based on an Ensemble Multilayer Perceptron (EMLP) model to give a proper correction. This correction is applied a posteriori to the SLAM estimation to compensate for the errors.
Moreover, the environment size, which is relatively high, cannot be dictated or limited a priori. Hence, we present a map management technique that is dedicated to 2D gridbased SLAM approaches; such a method ensures seamless navigation with stable resource requirements (i.e. memory and processor load) independently of the size of the environment and the length of the journey.
The approaches presented in this thesis are demonstrated and validated with a series of investigations with an extensive experimental evaluation carried out on open datasets and on our real vehicular platforms under real-life constraints.