A Vision based Terrain Classification Interface for Walking Machines

As walking machines have big advantages in moving through rough and undiscovered terrains compared to wheeled vehicles, a new problem raised: different terrains require different gaits. Thus, terrain classification plays an important role in autonomous outdoor navigation. Since computer vision is a rapidly growing field and inspired by biology, this thesis presents a vision based terrain classification interface for walking machines. For this purpose, only a single camera is required to capture terrain surfaces. Walking machines equipped with this interface are enabled to adjust their gait according to the terrain. Hence, a walking machine is able to walk efficiently in impassable areas. In order to design an online classifier which can run in realtime, different methods from image processing and machine learning fields are combined. The thesis provides both a color-based and a feature-based classification algorithm which are comprehensively examined to maximize the performance. The proposed classification algorithm is tested with a hexapod robot (AMOS II) which then becomes able to interact with the environment and thus change the gait patterns according to the classification outputs. Finally, the proposed algorithm is evaluated with a benchmark image set from the object classification field and with a huge terrain image set including five different classes. For the terrain images the algorithm yields an accuracy up to 98% at reasonable frame rate. Thus, the thesis provides a stable and robust terrain classification framework which can be used for various walking machines. Acknowledgements In the process of developing this thesis, many people were involved directly and indirectly, whom I want to express my thanks. To begin with, I want to thank my advisor Prof. Dr. Wörgötter for his friendly care of the thesis. I also want to thank Prof. Dr. May for agreeing to take the role of second reviewer. I thank Eren Erdal Aksoy who opened many doors for me. Everytime I was welcome and I could ask any questions which often led to extensible discussions. He spent a lot of time by proofreading this thesis and suggested me many valuable hints. Last but not least I want to thank Dr.-Ing. Poramate Manoonpong for inspiring the subject of this thesis and providing the AMOS II robot on which demonstrative experiments could be made.