The aim of this study is to analyze, compare and retain the most relevant tracking methods likely to respect the constraints of embedded systems, such as Micro Aerial Vehicles (MAVs), Unmanned Aerial Vehicles (UAVs) and intelligent glasses, in order to find a new robust embedded tracking system. A typical VINS consists of a monocular camera that provides visual data (frames), and a low-cost Inertial Measurement Unit (IMU), a Micro-Electro-Mechanical System (MEMS) that measures inertial data. This combination is very successful in the system navigation field thanks to the advantages that the sesensors provide, mainly in terms of accuracy, cost and reactivity. Over thelast decade, various sufficiently accurate tracking algorithms and Visual Inertial Navigation Systems (VINS) have been developed, however, they require greater computational resources. In contrast, embedded systems are characterized by their high integration constraints and limited resources. Thus,in this thesis, a solution for embedded architecture, relaying on efficient algorithms and providing less computational load, is proposed.First, relevant tracking algorithms are studied focusing on their accuracy, robustness, and computational complexity. In parallel, numerous recent embedded tracking computation architectures are also discussed. Then, our robust visual-inertial tracking approach, called : "Context Adaptive Visual Inertial SLAM", is introduced. It alternates between visual KLT-ORB and EKF Visual-Inertial tracking, according to the navigation context, thanks to the proposed execution control module. The latter uses several parameters concerning the scene’s appearance, the motion types, etc. Consequently, tracking continuity, robustness and accuracy are improved, even in difficult conditions. Moreover, our proposal is suited to embedded systems integration, given the low algorithms computational complexity and the implemented PoIs management leading to decrease the number of PoIs as well as the occurrences of their detection. All our experiments and tests was performed using the different EuRoC dataset sequences.