One of the more important tasks to be executed on a mobile robot, concerns the detection of obstacles during the robot motions. Many methods have been proposed for this function: nevertheless their performances are limited when applied in a structured environment made highly dynamic and cluttered due to humans. This document presents a visual and flexible system for obstacle detection in such an environment. The system is made of several micro-cameras fixed all around the robot body, and of a programmable electronic board. The camera number must be large enough (4 in the current version, 8 in the future one), so that real-time performances mandatory for such a function, cannot be reached from a standard multipurpose processor. It makes compulsory to design and to implement a hardware architecture devoted for image processing. The execution of parallel processes on FPGAs allows to reach real-time performances, while minimizing the required energy and the system cost. The system objective consists in building and updating a robot-centered occupancy grid while the robot is navigating. This function must be executed at 30Hz, in order to minimize the latency between image acquisition and obstacle detection. The detection of occupied ground areas is given by a classification algorithm, using an AdaBoost classifier on characteristic vectors. These vectors are built from color and texture attributes. For the color, the CIE-Lab space has been selected because it allows a better invariance according to the light variations. For the texture, an original method has been proposed adapting the Unser approach based on sum and difference histograms. This approach has been modified in order to reduce significantly the resources required to compute the texture attributes, while providing a fine model for every object detected on a scene acquired by each micro-camera. Each pixel in every image is classified as Ground or Obstacle, with respect to its color and texture attributes. Once a pixel is classified, it is projected on the ground plane in order to update the current occupancy grid built to represent the environment. Many parameters for our approach have been selected in order to develop a system with the better trade-off between performances and consumed resources. Every proposed architecture is evaluated using curves between classification performances and required resources. These architectures have been developed in VHDL using the Altera tool boxes; this classical approach has been compared with a method based on tools providing high level synthesis (Gaut, labview...). Finally all architectures avec been implemented and evaluated on a Stratix3 development kit connected to four cameras, and embedded on a mobile robot.
