Section: New Results

Scene Understanding with Computer Vision

Participants : Maximilian Jaritz, Raoul de Charette, Rafael Colmenares, Ziyang Hong, Fawzi Nashashibi.

This axis is in the continuation of previous year axis on scene understanding. It is crucial for autonomous driving. While last year we focused more on road estimation and ego velocity estimation (research report [51]), this year we focused on object recognition either from a single RGB camera or from a fusion of sensors (PhD of Maximilian Jaritz). Road estimation was also extended using graph energy minimization techniques and lead to interesting results in the scope of Rafael Colmenares internship. For object recognition a number of popular deep learning techniques were evaluated and the outcome of this evaluation study is that existing approaches suffers either from performance issues or processing time issues. In the scope of Maximilian Jaritz thesis a multi-modal approach is being developed were RGB and LiDAR are used to detect objects in the direct vicinity of the autonomous car. Preliminary results using state-of-the-art network architecture leads to satisfactory performances in terms of precision but a non-optimal localization of their spatial position (especially when rotated).