Efficient Stream-Based Active Learning Initialization for Legged Robots Based on a PCA/K-Means Image Selection Approach

In recent years walking robots have become a promising tool in exploration and search and rescue missions in partially unknown and changing environments. When utilizing machine learning models for segmentation and classification unknown environments can decrease the performance of the deployed models, as they might encounter objects for which they don’t have sufficient training data. Active Learning can be used to select new data to be annotated during the mission and provide the model with the required knowledge as long as data that is useful to the model can be identified. Shortly after the start of the robot mission, this might not be the case and data might be selected randomly, leading to worse performance of the model or outright failure of the training process. In this work, we both discuss the initialization problem and propose an approach to use image set analysis algorithms on the unlabeled set of data the robot would have access to find key images that can be utilized to successfully start a Stream-Based Active Learning process. We evaluate our approach on a small set of images as well as multiple data streams to show the validity of the approach itself as well as important metrics in identifying the right time to transition from the initialization to the main learning process.