Contributions to vehicle recognition in occupancy grids for an embedded perception system ; Contributions à la reconnaissance de véhicules dans des grilles d’occupation pour un système de perception embarqué

This PhD thesis manuscript presents the work carried out over a three-year period on the subject of recognizing vehicles present in the scene from the two-dimensional occupancy grid environment model. This subject takes place in the autonomous vehicle context, and represents an important step in the perception system enabling a vehicle to create an internal representation of its environment, which is then used to make appropriate navigation decisions. The constraint inherent to autonomous vehicles of real-time execution of the algorithms developed is specifically addressed by the search for the most economical methods in terms of computational and memory costs.Three main contributions are presented in this manuscript. The first is the creation of a dataset of occupancy grids annotated with the positions and dimensions of the vehicles present in the scene. This first contribution is a prerequisite for the deep learning methods for vehicle recognition proposed in the second and third contributions. The second contribution is a study of different convolutional neural network models inspired by state of the art models for image-based object detection to detect vehicles present in occupancy grids via oriented bounding boxes prediction. The final contribution is the proposal of a new approach for reducing the dimensionality of occupancy grids via their transformation to the frequency domain, and the proposal of a neural architecture for segmenting vehicles present in occupancy grids from this low-dimensional representation.The different vehicle recognition methods proposed are evaluated according to the quality of the predicted results and the computational costs of prediction, which are materialized by the prediction time per occupancy grid and the memory cost of each prediction. The first approach, inspired by convolutional neural networks used for object detection on images, validates the use of occupancy grids as a sufficiently rich environment model to enable recognition of vehicles present in the environment. ...