PhD co-funded by EURECOM and IMT Lille-Douai
Accidents involving Vulnerable Road Users remain a significant issue for road safety, accounting for more than 25% of road fatalities in the European Union (World Health Organization, 2015). To adress this problem a solution is to develop the next generation of active safety systems for protecting Vulnerable Road Users (VRU) able of improved performance by expanding the scope of scenarios for a better understanding of vehicle-VRU accidents i.e. the relationship between behaviour and conflicting situations. In achieving this goal, researchers propose an early and in-depth understanding of how vehicles and VRUs interact in real traffic situations by analysing accident (or near accident) events from video suveillance databases.
Traffic video observations are usually used to evaluate road safety of the infrastructure. Some protocols use several human observers to identify VRU/vehicule conflicts and to deeply describe the interactions. The protocol is based on the human capabilities to analyse accuratly a high number of situations and to determine the variables which the critical level depends on. This human task is time consumming. It can be improved thanks to the recent advances in image processing and modelling tools that can faster provide quite accurate information from big video databases.
Mobile objects detection, classification (pedestrian, moped, bycicle, car, bus, truck etc.) and trajectography estimation are three image processing functions particularly sought to qualify a conflict and to tune a system able to better mitigate the situation. Pedestrians head position or cyclists arm movements are indicators that also signal VRU intent in the interaction with the vehicle. Both last information can be extracted from de video sequences if the camera is near the interaction and offers a detailed image of the pedestrians or the cyclist (on-side road cameras).
It is very difficult to define a critical level function from the extracted information, especially if this function has to be valid for different environment. A solution is to estimate this decision function using learning and modelling techniques applied to big video datasets. Many techniques based on a labelled learning datasets are well known and described in the literature. To reduce the constraint of a hand-labelling task, incremental learning has been proposed. Recently Deep Learning solutions are very promising in replacing classical methods (SVM, GMM, HMM) for supervised and unsupervised modelling.
The literature is full of algorithms that can perform these image processing and learning task. But often it is not possible to guarantee the expected accuracy and robustness more specifically to deal with large video databases. In this PhD subject, we propose to improve existing algorithms for the following objectives:
Because of mutations in urban mobility and the increasing number of cyclists and more generally green mobility devices (moped, motorized scooter...), this PhD subject will focus specially on urban environments, where all green mobility devices share majority of fatalities. The candidate will construct and evaluate her/his work on large video databases acquired on several urban sites.
image processing, machine learning
Computer tools: C, C++, matlab, OpenCV
Sébastien Ambellouis (IMT/IFSTTAR), Jacques Bonnaert (IMT), Pr. Jean-Luc Dugelay (Eurecom)
(c) GdR 720 ISIS - CNRS - 2011-2015.