École Nationale Supérieure de
Techniques Avancées

Résumé

This research internship is the first part of a larger project dedicated to intelligent vehicles whose final goal is to calculate and share trajectories among autonomous vehicles. With such cooperation between vehicles comes an increase in safety by enabling shared detection of obstacles. Consequently, those not directly visible to a given vehicle will now be detected too. The work carried out consisted of merging data from different sensors, especially Li- DAR and camera. Ultimately, the algorithms developed could precisely identify obstacles in the nearby environment. Initially, a bird’s eye view (BEV) representation was constructed from the LiDAR point clouds, allowing for a usable 2D visualization. A detection algorithm was then im- plemented to generate bounding boxes around the detected objects. These boxes were then projected onto the camera images using a series of geometric transformations based on the calibration matrices of the KITTI dataset, a dataset provided by Karlsruhe Insti- tute of Technology (KIT) and Toyota Technological Institute, that was used throughout the internship to test the different algorithms. Particular attention was paid to the quality of the visual overlay by correcting shifts due to calibration errors. In parallel, ground filtering using RANSAC allowed for the isolation of relevant obstacles. This work forms an essential foundation for the next step of the project, which is the dynamic estimation of trajectories. It has allowed for the exploration of concrete issues related to computer vision, projective geometry, and sensor fusion in a demanding appli- cation framework