In automotive industry, industrial robots are widely used in production lines for many tasks such as welding, painting or assembly. Their use requires, from users, both a good manipulation and robot control. Recently, new tools have been developed to realize fast and accurate trajectories in many production sectors by using the real prototype of vehicle or a generalized design within a virtual simulation platform. However, many issues could be considered in these cases: the delay between the design of the vehicle and its production is often important, moreover, the virtual modeling presents a non realistic aspect of the real robot and vehicle, so this factor could introduce localization inacurracies in performing trajectories. Our work is registered as a part of TRI project (Teleteaching Industrial Robots) which aims to realize a demonstrator showing the interaction of industrial robots with virtual components and allowing to train users to perform successfully their tasks on a virtual representation of a production entity. In this project we make use of Augmented Reality (AR) techniques to overlay virtual objects onto the real world in order to enhance the user's perception and interaction while performing a specific industrial task. The idea is to allow the real robot to teach trajectories of an automotive task thanks to vehicle virtual model. The pose accuracy is prerequisite of our application since it allows a reliable teaching of the real trajectory. Therefore, we survey some vision-based pose computation algorithms and present a method that offers increased robustness and accuracy in the context of real-time AR tracking. Our aim is to determine the performance of these pose estimation methods in term of errors and distance evaluation. The evaluation of the pose estimation methods was obtained using a series of tests and an experimental protocol. The analysis of results shows the performance of algorithms in term of accuracy, stability and convergence.