Skip to Main content Skip to Navigation

Estimation du contexte par vision embarquée et schémas de commande pour l’automobile

Abstract : To take relevant decisions, autonomous systems have to continuously estimate their environment via embedded sensors. In the case of 'intelligent' vehicles, the estimation of the context focuses both on objects perfectly known such as road signs (vertical or horizontal), and on objects unknown or difficult to describe due to their number and variety (pedestrians, cyclists, other vehicles, animals, any obstacles on the road, etc.). Now, the a contrario modelling provides a formal framework adapted to the problem of detection of variable objects, by modeling the noise rather than the objects to detect. Our main contribution in this PhD work was to adapt the probabilistic NFA (Number of False Alarms) measurements to the problem of detection of objects simply defined either as having an own motion, or salient to the road plane. A highlight of the proposed algorithms is that they are free from any detection parameter, in particular threshold. A first NFA criterion allows the identification of the sub-domain of the image (not necessarily connected pixels) whose gray level values are the most amazing under Gaussian noise assumption (naive model). A second NFA criterion allows then identifying the subset of maximum significant windows under binomial hypothesis (naive model). We prove that these measurements (NFA) can also be used for the estimation of intrinsec parameters, for instance either the 6D movement of the onboard camera, or a binarisation threshold. Finally, we prove that the proposed algorithms are generic and can be applied to different kinds of input images, for instance either radiometric images or disparity maps. Conversely to the a contrario approach, the Markov models allow to inject a priori knowledge about the objects sought. We use it in the case of the road marking classification. From the context estimation (road signs, detected objects), the control part includes firstly a specification of the possible trajectories and secondly the laws to achieve the selected path. The possible trajectories are grouped into a bundle, and various parameters are used to set the local geometric invariants (slope, curvature). These parameters depend on the vehicle context (presence of vulnerables, fixed obstacles, speed limits, etc ... ), and allows determining the selected the trajectory from the bundle. Differentially flat system is indeed fully parameterized by its flat outputs and their derivatives. Another feature of this kind of systems is to be accurately linearized by endogenous dynamics feed-back and coordinate transformation. Tracking stabilizer is then trivially obtained from the linearized system.
Complete list of metadata

Cited literature [138 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Monday, April 22, 2013 - 10:29:41 AM
Last modification on : Wednesday, October 14, 2020 - 3:57:11 AM
Long-term archiving on: : Tuesday, July 23, 2013 - 4:12:20 AM


Version validated by the jury (STAR)


  • HAL Id : tel-00811797, version 1



Moez Ammar. Estimation du contexte par vision embarquée et schémas de commande pour l’automobile. Autre [cond-mat.other]. Université Paris Sud - Paris XI, 2012. Français. ⟨NNT : 2012PA112425⟩. ⟨tel-00811797⟩



Record views


Files downloads