Object-relative mobile robot navigation is essential for a variety of tasks, e.g. autonomous critical infrastructure inspection, but requires the capability to extract semantic information about the objects of interest from raw sensory data. While deep learning-based (DL) methods excel at inferring semantic object information from images, such as class and relative 6 degree of freedom (6-DoF) pose, they are computationally demanding and thus often not suitable for payload constrained mobile robots. In this letter we present a real-time capable unmanned aerial vehicle (UAV) system for object-relative, closed-loop navigation with a minimal sensor configuration consisting of an inertial measurement unit (IMU) and RGB camera. Utilizing a DL-based object pose estimator, solely trained on synthetic data and optimized for companion board deployment, the object-relative pose measurements are fused with the IMU data to perform object-relative localization. We conduct multiple real-world experiments to validate the performance of our system for the challenging use case of power pole inspection. An example closed-loop flight is presented in the supplementary video.

The capability to extract task specific, semantic information from raw sensory data is a crucial requirement for many applications of mobile robotics. Autonomous inspection of critical infrastructure with Unmanned Aerial Vehicles (UAVs), for example, requires precise navigation relative to the structure that is to be inspected. Recently, Artificial Intelligence (AI)-based methods have been shown to excel at extracting semantic information such as 6 degree-of-freedom (6-DoF) poses of objects from images. In this paper, we propose a method combining a state-of-the-art AI-based pose estimator for objects in camera images with data from an inertial measurement unit (IMU) for 6-DoF multi-object relative state estimation of a mobile robot. The AI-based pose estimator detects multiple objects of interest in camera images along with their relative poses. These measurements are fused with IMU data in a state-of-the-art sensor fusion framework. We illustrate the feasibility of our proposed method with real world experiments for different trajectories and number of arbitrarily placed objects. We show that the results can be reliably reproduced due to the self-calibrating capabilities of our approach.

We present a novel approach for training deep neural networks in a Bayesian way. Classical, i.e. non-Bayesian, deep learning has two major drawbacks both originating from the fact that network parameters are considered to be deterministic. First, model uncertainty cannot be measured thus limiting the use of deep learning in many fields of application and second, training of deep neural networks is often hampered by overfitting. The proposed approach uses variational inference to approximate the intractable a posteriori distribution on basis of a normal prior. The variational density is designed in such a way that the a posteriori uncertainty of the network parameters is represented per network layer and depending on the estimated parameter expectation values. This way, only a few additional parameters need to be optimized compared to a non-Bayesian network. We apply this Bayesian approach to train and test the LeNet architecture on the MNIST dataset. Compared to classical deep learning, the test error is reduced by 15%. In addition, the trained model contains information about the parameter uncertainty in each layer. We show that this information can be used to calculate credible intervals for the prediction and to optimize the network architecture for a given training data set.