Driver Assistance by Asynchronous Camera Ring – DrAACaR

This project deals with the development of an asynchronous hybrid network composed of wide-angle cameras and 2D lasers in the context of Driver Assistance System (DAS). It is well established that optical and range sensors are very useful for tasks such as lane keeping assistant, road detection, motion estimation, obstacle avoidance, … Many different configurations are possible according to the optical devices (perspective, catadioptric, fish-eye cameras, IR cameras, …) and of range sensors (visible/invisible, 2D/3D) and their numbers (mono, stereo, multiple). In this project, we are mainly interesting by multiple wide-angle fisheye cameras in collaboration with a set of 2D lasers. This configuration has the main advantage of being low-cost but also very effective. Generally, these multi-sensor systems are composed of devices that are perfectly synchronized and calibrated. However, here we do not impose any constraints about the synchronization of these cameras and lasers as well as about their relative position. These aspects of unsynchronization and unknown calibration constitute the main contributions of our project since it permits to obtain a low cost system easy to implement on a car (no synchronization system and a limited wiring). Moreover they represent a real scientific challenge since several fundamental issues have to be solved.
The different applications that we propose to develop for DAS are the Dense Hybrid Structure from Motion around the vehicle, road detection for vehicle localization, obstacle detection.
The Dense Hybrid Structure from Motion takes as inputs a set of images and 3-D data and aims to estimate the 3D structure of the scene and the vehicle location. Contrary to existing methods, we propose to combine the laser scanners with the omnidirectional fisheye cameras in a so-called Hybrid Structure from Motion, while facing the important and challenging unsynchronization issue.
Road detection consists in extracting the drivable path of the scene in the acquired data (images, laser, etc…). Obstacle detection/avoidance refers to the tasks of, first, detecting the objects (cars, pedestrians, etc…) that could collide with the vehicle, and secondly, modifying the vehicle trajectory to avoid the collision. Therefore developing robust obstacle detection methods is a crucial task to enhance the security and can save thousands of lives per year. The proposed omnidirectional vision-laser system is designed to observe and detect the potential obstacles in all the directions. Our method consists in merging the images and the laser scan 3-D data obtained by the fisheye cameras and the laser scanners to detect and track the dangerous obstacles that could collide with the vehicle.
A common issue to success in these three tasks is the need of methods able to match together the sequences obtained from the different cameras and laser scanners. However, these sequences being not synchronized, this matching process becomes particularly critical and the existing methods have a lot of difficulties. The aim will then consist in developing new matching techniques while dealing with the un-synchronization. Some interesting characteristics are the following: the cameras and the laser scanners have an unknown but rigid calibration, ground vehicles drive on a (locally) flat ground and verify the non-holonomic constraints, and urban scenes contain structured features.
We will develop an autonomous vehicle testbed that our Korean partners have been working on since 2009. A commercial SUV vehicle will be provided by the Hyundai Motors Company and will be modified to accept CAN-based command for an automatic control of the vehicle. It will be upgraded to provide sufficient DC and AC power for various onboard components such as navigation sensors, lasers, cameras, and computers.