Embedded vision-based SLAM: A model-driven approach

This paper presents a model driven design approach for embedded heterogeneous systems. This method relies on the use of a Data-Flow model to describe an application structure and behavior. This model is then transformed through an iterative process to refine an execution model using hardware/software resources described by the user. This approach is composed of a set of tasks and their matching validation to ensure tracking of the implementation validity and performance along the design flow. This method is then used for the implementation of a full vision-based Simultaneous Localization And Mapping systems on an heterogeneous architecture. The SLAM algorithm aims at computing the position of a mobile and build a map of its surrounding using a set of sensors. In vision-based SLAM this position is estimated by tracking features along time in images acquired by a set of cameras. The algorithm relies on several compute-intensive tasks for both vision and estimation making this problem hard to solve on resources limited platform such as embedded systems.