Abstract:
This thesis is concerned with the robustness of robots’ spatial cognition. Two separateissuesareconsidered: Illuminationinvarianceandplacerecognition. Illumination invariance ensures that appearances of the same scene under different illumination conditions do not change. Four existing approaches are considered as well as a new approach is proposed in this framework. Aforementioned two issues are addressed in two stages individually. In the first stage, appearances collected from robot’s image sensor are reconstructed in order to make robot interpret surrounding environment robust to illumination variations by using any descriptor. Illumination invariant appearances are employed so that robot is able to both distinguish one place from another in a single environment and recognize any place regardless of the change in the illumination conditions. An extensive comparative experimental evaluation serves to demonstrate how each method performs with respect to appearance similarity, place detection and place recognition. On the other hand, the second stage focuses on robot’s spatial memory. Simple yet efficient method is proposed in order to address the place learning and recognition performance issues. Detected places are clustered into smaller place clusters called “subplaces” where each subplace includes canonical appearances associated with the corresponding place. Thus, canonical appearances are evaluated for learning andrecognitioninsteadofemployingthewholeplace. Itguaranteesthatonlydominant appearances in a place are recognized and thus achieves a better performance by the help of these canonical views. Likewise in the first stage, comparative experiments are conducted on various datasets and on the robot in real time with and without subplace approach. Learning and recognition performance alterations are discussed in detail.
