Visual odometry is an essential problem for mobile robots. Traditional methods for solving VO mostly utilize geometric optimization. While capable of achieving high accuracy, these methods require accurate sensor calibration and complicated parameter tuning to work well in practice. With the rise of deep learning, there has been increased interest in the end-to-end, learning-based methods for VO, which have the potential to improve robustness. However, learning-based methods for VO so far are less accurate than geometric methods. We argue that one of the main issues is that the current ego-motion estimation task is different from other problems where deep learning has been successful such as object detection. We define a novel cost function for learning-based VO considering the mathematical properties of the group homomorphism. In addition to the standard L2 loss, we incorporate losses based on the identity, inverse and closure properties of SE(3) rigid motion. Furthermore, we propose to reduce the VO drift by estimating the drivable regions using semantic segmentation and incorporate this information into a pose graph optimization. Experiments on KITTI datasets show that the novel cost function can improve ego-motion estimation compared to the state-of-the-art and the drivable region-based correction further reduces the VO drift.

Recently, reinforcement learning with deep neural networks has achieved great success in challenging continuous control problems such as 3D locomotion and robotic manipulation. However, in real-world control problems, the actions one can take are bounded by physical constraints, which introduces a bias when the standard Gaussian distribution is used as the stochastic policy. In this work, we propose to use the Beta distribution as an alternative and analyze the bias and variance of the policy gradients of both policies. We show that the Beta policy is bias-free and provides significantly faster convergence and higher scores over the Gaussian policy when both are used with trust region policy optimization (TRPO) and actor critic with experience replay (ACER), the state-of-the-art on- and off-policy stochastic methods respectively, on OpenAI Gym’s and MuJoCo’s continuous control environments.

Wire detection is a key capability for safe navigation of autonomous aerial vehicles and is a challenging problem as wires are generally only a few pixels wide, can appear at any orientation and location, and are hard to distinguish from other similar looking lines and edges. We leverage the recent advances in deep learning by treating wire detection as a semantic segmentation task, and investigate the effectiveness of convolutional neural networks for the same. To find an optimal model in terms of detection accuracy and real time performance on a portable GPU, we perform a grid search over a finite space of architectures. Further, to combat the issue of unavailability of a large public dataset with annotations, we render synthetic wires using a ray tracing engine, and overlay them on 67K images from flight videos available on the internet. We use this synthetic dataset for pretraining our models before finetuning on real data, and show that synthetic data alone can lead to pretty accurate detections qualitatively as well. We also verify if providing explicit information about local evidence of wiry-ness in the form of edge and line detection results from a traditional computer vision method, as additional channels to the network input, makes the task easier or not. We evaluate our best models from the grid search on a publicly available dataset and show that they outperform previous work using traditional computer vision and various deep net baselines of FCNs, SegNet and E-Net, on both standard edge detection metrics and inference speed. Our top models run at more than 3Hz on the NVIDIA Jetson TX2 with input resolution of 480×640, with an Average Precision score of 0.73 on our test split of the USF dataset.

The last decade has seen a massive growth in applications for Micro-Aerial Vehicles (MAVs), due in large part to their versatility for data gathering with cameras, LiDAR and various other sensors. Their ability to quickly go from assessing large spaces from a high vantage points to flying in close to capture high-resolution data makes them invaluable for applications where we are interested in a specific target with an a priori unknown location, e.g. survivors in disaster response scenarios, vehicles in surveillance, animals in wildlife monitoring, etc., a task we will refer to scouting. Our ultimate goal is to enable MAVs to perform autonomous scouting. In this paper, we describe a semantic mapping system designed to support this goal. The system maintains a 2.5D map describing its belief about the location of semantic classes of interest, using forward-looking cameras and state estimation. The map is continuously updated on the fly, using only onboard processing. The system couples a deep learning 2D semantic segmentation algorithm with a novel mapping method to project and aggregate the 2D semantic measurements into a global 2.5D grid map. We train and evaluate our segmentation method on a novel dataset of cars labelled in oblique aerial imagery. We also study the performance of the mapping system in isolation. Finally, we show the integrated system performing a fully autonomous car scouting mission in the field.

We consider the problem of understanding the 3D layout of indoor corridor scenes from a single image in real time. Identifying obstacles such as walls is essential for robot navigation, but also challenging due to the diversity in structure, appearance and illumination of real-world corridor scenes. Many current single-image methods make Manhattan-world assumptions, and break down in environments that do not meet this mold. They also may require complicated hand-designed features for image segmentation or clear boundaries to form certain building models. In addition, most cannot run in real time. In this paper, we propose to combine machine learning with geometric modelling to build a simplified 3D model from a single image. We first employ a supervised Convolutional Neural Network (CNN) to provide a dense, but coarse, geometric class labelling of the scene. We then refine this labelling with a fully connected Conditional Random Field (CRF). Finally, we fit line segments along wall-ground boundaries and “pop up” a 3D model using geometric constraints. We assemble a dataset of 967 labelled corridor images. Our experiments on this dataset and another publicly available dataset show our method outperforms other single image scene understanding methods in pixelwise accuracy while labelling images at over 15Hz.

We present a system for the detection of small and potentially obscured obstacles in vegetated terrain. The key novelty of this system is the coupling of a volumetric occupancy map with a 3D Convolutional Neural Network (CNN), which to the best of our knowledge has not been previously done. This architecture allows us to train an extremely efficient and highly accurate system for detection tasks from raw occupancy data. We apply this method to the problem of detecting safe landing zones for autonomous helicopters from LiDAR point clouds. Current methods for this problem rely on heuristic rules and use simple geometric features. These heuristics break down in the presence of low vegetation, as they do not distinguish between vegetation that may be landed on and solid objects that should be avoided. We evaluate the system with a combination of real and synthetic range data. We show our system outperforms various benchmarks, including a system integrating various hand-crafted point cloud features from the literature.

Robust object recognition is a crucial skill for robots operating autonomously in real world environments. Range sensors such as LiDAR and RGBD cameras are increasingly found in modern robotic systems, providing a rich source of 3D information that can aid in this task. However, many current systems do not fully utilize this information and have trouble efficiently dealing with large amounts of point cloud data. In this paper, we propose VoxNet, an architecture to tackle this problem by integrating a volumetric Occupancy Grid representation with a supervised 3D Convolutional Neural Network (3D CNN). We evaluate our approach on publicly available benchmarks using LiDAR, RGBD, and CAD data. VoxNet achieves accuracy beyond the state of the art while labeling hundreds of instances per second.

This paper poses object category detection in images as a type of 2D-to-3D alignment problem, utilizing the large quantities of 3D CAD models that have been made publicly available online. Using the "chair" class as a running example, we propose an exemplar-based 3D category representation, which can explicitly model chairs of different styles as well as the large variation in viewpoint. We develop an approach to establish part-based correspondences between 3D CAD models and real photographs. This is achieved by (i) representing each 3D model using a set of view-dependent mid-level visual elements learned from synthesized views in a discriminative fashion, (ii) carefully calibrating the individual element detectors on a common dataset of negative images, and (iii) matching visual elements to the test image allowing for small mutual deformations but preserving the viewpoint and style constraints. We demonstrate the ability of our system to align 3D models with 2D objects in the challenging PASCAL VOC images, which depict a wide variety of chairs in complex scenes.

Histograms of Local Binary Patterns (LBPs) and variations thereof are a popular local visual descriptor for face recognition. So far, most variations of LBP are designed by hand or are learned with non-supervised methods. In this work we propose a simple method to learn discriminative LBPs in a supervised manner. The method represents an LBP-like descriptor as a set of pixel comparisons within a neighborhood and heuristically seeks for a set of pixel comparisons so as to maximize a Fisher separability criterion for the resulting histograms. Tests on standard face recognition datasets show that this method can create compact yet discriminative descriptors.