In the field of obstacle detection and tracking for vision-based ADAS (Advanced Driver Assistance System), it is necessary to perform short-term vehicle localisation. Vision based SLAM (Simultaneous Localization and Mapping) solves this problem by combining the vehicle state estimation (local pose and speeds) and an incremental modelling of the environment. The environment is perceived by extracting features (interest points) in a sequence of images and tracking them over time to allow an incremental landmarks map construction. The perception step leads to an important computational load which affects very significantly the system latency and throughput. Co-design methodologies allow to design a mixed processing architecture optimized for a...