On-the-fly detection of novel objects in indoor environments

Many robotic applications require the detection of new objects in known environments. Common approaches navigate in the environment using pre-defined waypoints and segment the scene at these waypoints. Without knowing where to find new objects, this process can be time-consuming and prone to detecting false positives. To overcome these limitations we propose an approach that combines navigation and attention in order to detect novel objects rapidly. We exploit the octomap, created by the robot while it navigates in the environment, as a pre-attention filter to suggest potential regions of interest. These regions are then visited to obtain a close-up view for better object detection and recognition. We evaluate our approach in a simulated as well as a real environment. The experiments show that our approach outperforms previous approaches in terms of runtime and the number of segmentation actions required to find all novel objects in the environment.