Implementation of Pure and Hybrid Visual Odometries on electric self-driving all-terrain vehicles

The design of a fully autonomous mobile platform is tightly related to the development of a reliable navigation system. In such context, the problem of the self-localisation covers one of the most important challenges. Nowadays both autonomous and non-autonomous vehicles mainly rely on GPS when dealing with pose estimation. As a drawback, the GPS system is likely to return a non-informative estimate for short-range displacements (i.e. indoor mobile robots operations), as well as in some places the signal can be either unreliable or totally absent. For the reasons mentioned above, a good self-contained localisation system needs to be integrated, in order to fulfill the localisation issue in a wider range of environments. The purpose of this research work is to develop a real-time application in order to get a vision-based pose estimate of a Polaris electric all-terrain vehicle equipped with an omnidirectional catadioptric camera. After a general overview of the state-of-the-art about visual odometry (VO), two approaches are compared: the former is a pure VO technique based on the hypothesis of ground planarity, whereas the latter is a hybrid approach which makes use of the speed information to offset the camera non-idealities. In the end, performances are evaluated and the best solution undergoes to code generation. The goal to pursue is to obtain an execution time at least lower than the camera frame rate, in order to obtain a feasible solution for a real-time application.