Active Learning For Outdoor Obstacle Detection

Abstract

Abstract-Real-world applications of mobile robotics call for increased autonomy, requiring reliable perception systems. Since manually tuned perception algorithms are difficult to adapt to new operating environments, systems based on supervised learning are necessary for future progress in autonomous navigation.Data labeling is a major concern when supervised learning is applied to the large-scale problems occuring in realistic robotics applications. We believe that algorithms for automatically selecting important data for labeling are necessary, and propose to employ active learning techniques to reduce the amount of labeling required to learn from a data set.In this paper we show that several standard active learning algorithms can be adapted to meet specific constraints characteristic to our domain, such as the need to learn from data with severely unbalanced class priors. We validate the solutions we propose by extensive experimentation on multiple realistic data sets captured with a robotic vehicle. Based on our results for the task of obstacle detection, we conclude that active learning techniques are applicable to our domain, and they can lead to significant reductions in the labeling effort required to use supervised learning in outdoor perception.

Active learning from demonstration for robust autonomous navigation

“…Novelty detection in and of itself is a rich field of research [8], and has seen many applications to mobile robotics [9], [10]. [11], [12] specifically propose density estimation in the context of active learning for robotic perception systems. Specifically, a large dataset consisting of a robot's recorded perceptual history is analyzed to identify the most unlikely single percepts (given the entire history).…”

“…However, there is another approach that can work with general cost functions. The Query by Bagging approach [12] combines the idea of Query by Committee [16], [17] with the idea of Bagging [18] to measure the uncertainty still inherent in a training set for a particular class of learner. The idea is to train multiple learners on different random subsets (with replacement) of the available training set, and see where they disagree.…”

An Information Theoretic Approach of Designing Sparse Kernel Adaptive Filters

“…One way is to use kernel density estimation as in [27] but this is problematic when the dimensionality of is high. Another way is to convert it to a parameter estimation problem by assuming a parametric distribution family.…”

“…In the case of unknown desired signal, we simply average over the posterior distribution of . By using (22), one has Neglecting the constant terms yields (27) Furthermore, under a memoryless uniform input assumption, it is simplified as (28) Therefore, ALD and variance criterion are a special case of the surprise criterion.…”

“…However, its application to active data selection is quite novel. The work of Dima and Hebert [27] is similar to ours but they used kernel density estimation while we use GP theory to derive an analytical solution. Defining the instantaneous learnable information contained on a data sample allows us to discard or include new exemplars systematically and curb the complexity of online kernel algorithms.…”