No abstract

Visual active tracking is a growing research topic in robotics due to its key role in applications such as human assistance, disaster recovery, and surveillance. In contrast to passive tracking, active tracking approaches combine vision and control capabilities to detect and actively track the target. Most of the work in this area focuses on ground robots, while the very few contributions on aerial platforms still pose important design constraints that limit their applicability. To overcome these limitations, in this paper we propose D-VAT, a novel end-to-end visual active tracking methodology based on deep reinforcement learning that is tailored to micro aerial vehicle platforms. The D-VAT agent computes the vehicle thrust and angular velocity commands needed to track the target by directly processing monocular camera measurements. We show that the proposed approach allows for precise and collision-free tracking operations, outperforming different state-of-the-art baselines on simulated environments which differ significantly from those encountered during training.

In the last decades, ego-motion estimation or visual odometry (VO) has received a considerable amount of attention from the robotic research community, mainly due to its central importance in achieving robust localization and, as a consequence, autonomy. Different solutions have been explored, leading to a wide variety of approaches, mostly grounded on geometric methodologies and, more recently, on data-driven paradigms. To guide researchers and practitioners in choosing the best VO method, different benchmark studies have been published. However, the majority of them compare only a small subset of the most popular approaches and, usually, on specific data sets or configurations. In contrast, in this work, we aim to provide a complete and thorough study of the most popular and best-performing geometric and data-driven solutions for VO. In our investigation, we considered several scenarios and environments, comparing the estimation accuracies and the role of the hyper-parameters of the approaches selected, and analyzing the computational resources they require. Experiments and tests are performed on different data sets (both publicly available and self-collected) and two different computational boards. The experimental results show pros and cons of the tested approaches under different perspectives. The geometric simultaneous localization and mapping methods are confirmed to be the best performing, while data-driven approaches show robustness with respect to nonideal conditions present in more challenging scenarios.

Simultaneous localization and mapping (SLAM) is one of the cornerstones of autonomous navigation systems in robotics and the automotive industry. Visual SLAM (V-SLAM), which relies on image features, such as keypoints and descriptors to estimate the pose transformation between consecutive frames, is a highly efficient and effective approach for gathering environmental information. With the rise of representation learning, feature detectors based on deep neural networks (DNNs) have emerged as an alternative to handcrafted solutions. This work examines the integration of sparse learned features into a state-of-the-art SLAM framework and benchmarks handcrafted and learning-based approaches by comparing the two methods through in-depth experiments. Specifically, we replace the ORB detector and BRIEF descriptor of the ORBSLAM3 pipeline with those provided by Superpoint, a DNN model that jointly computes keypoints and descriptors. Experiments on three publicly available datasets from different application domains were conducted to evaluate the pose estimation performance and resource usage of both solutions.

No abstract