Real-time Stereovision by an Integrated Sensor

Abstract : This paper concerns the integration of a complex vision algorithm on a dedicated architecture. Perception is required on a vehicle (robot or car), especially in order to detect obstacles along a trajectory for a robot or on the road for a car. This function must satisfy very hard real-time constraints so that an avoidance strategy could be executed by the onboard system on a robot or by a driver on a car: it is required to detect obstacles at 100Hz. Several sensors have been used to acquire perceptual data from which obstacles are detected: we propose to use a stereovision sensor. This algorithm is so complex that it must be integrated on a dedicated architecture to provide a high resolution 3D reconstruction of the environment at 100Hz. At first, this paper describes a correlation-based stereovision algorithm, and a generic architecture including several FPGA-based boards, designed to execute visual functions at very high frequency. Then we propose a strategy to implement our stereovision algorithm on this computing unit, with four FPGA-based boards.