With the Autonomous Vehicle (AV) industry shifting towards machine-learned approaches for motion plan-ning [1], the performance of self-driving systems is starting to rely heavily on large quantities of expert driving demon-strations. However, collecting this demonstration data typically involves expensive HD sensor suites (LiDAR + RADAR + cameras), which quickly becomes financially infeasible at the scales required. This motivates the use of commodity sensors like cameras for data collection, which are an order of mag-nitude cheaper than HD sensor suites, but offer lower fidelity. Leveraging these sensors for training an AV motion planner opens a financially viable path to observe the ‘long tail’ of driving events. As our main contribution we show it is possible to train a high-performance motion planner using commodity vision data which outperforms planners trained on HD-sensor data for a fraction of the cost. To the best of our knowledge, we are the first to demonstrate this using real-world data. We compare the performance of the autonomy system on these two different sensor configurations, and show that we can compensate for the lower sensor fidelity by means of increased quantity: a planner trained on 100h of commodity vision data outperforms the one with 25h of expensive HD data (see Fig. 1). We also share the engineering challenges we had to tackle to make this work.

Localization is an integral part of reliable robot navigation, and long-term autonomy requires robustness against perceptional changes in the environment during localization. In the context of vision-based localization, such changes can be caused by illumination variations, occlusion, structural development, different weather conditions, and seasons. In this paper, we present a novel approach for localizing a robot over longer periods of time using only monocular image data. We propose a novel data association approach for matching streams of incoming images to an image sequence stored in a database. Our method exploits network flows to leverage sequential information to improve the localization performance and to maintain several possible trajectories hypotheses in parallel. To compare images, we consider a semidense image description based on histogram of oriented gradients features as well as global descriptors from deep convolutional neural networks trained on ImageNet for robust localization. We perform extensive evaluations on a variety of datasets and show that our approach outperforms existing state-of-the-art approaches.

Visual place recognition under difficult perceptual conditions remains a challenging problem due to changing weather conditions, illumination and seasons. Long-term visual navigation approaches for robot localization should be robust to these dynamics of the environment. Existing methods typically leverage feature descriptions of whole images or image regions from Deep Convolutional Neural Networks. Some approaches also exploit sequential information to alleviate the problem of spatially inconsistent and non-perfect image matches. In this paper, we propose a novel approach for learning a discriminative holistic image representation which exploits the image content to create a dense and salient scene description. These salient descriptions are learnt over a variety of datasets under large perceptual changes. Such an approach enables us to precisely segment the regions of an image which are geometrically stable over large time lags. We combine features from these salient regions and an off-the-shelf holistic representation to form a more robust scene descriptor. We also introduce a semantically labeled dataset which captures extreme perceptual and structural scene dynamics over the course of 3 years. We evaluated our approach with extensive experiments on data collected over several kilometers in Freiburg and show that our learnt image representation outperforms off-the-shelf features from the deep networks and hand-crafted features.

Precise localization of robots is imperative for their safe and autonomous navigation in both indoor and outdoor environments. In outdoor scenarios, the environment typically undergoes significant perceptual changes and requires robust methods for accurate localization. Monocular camera-based approaches provide an inexpensive solution to such challenging problems compared to 3D LiDAR-based methods. Recently, approaches have leveraged deep convolutional neural networks (CNNs) to perform place recognition and they turn out to outperform traditional handcrafted features under challenging perceptual conditions. In this paper, we propose an approach for directly regressing a 6-DoF camera pose using CNNs and a single monocular RGB image. We leverage the idea of transfer learning for training our network as this technique has shown to perform better when the number of training samples are not very high. Furthermore, we propose novel data augmentation in 3D space for additional pose coverage which leads to more accurate localization. In contrast to the traditional visual metric localization approaches, our resulting map size is constant with respect to the database. During localization, our approach has a constant time complexity of O(1) and is independent of the database size and runs in real-time at ~80 Hz using a single GPU. We show the localization accuracy of our approach on publicly available datasets and that it outperforms CNN-based state-of-the-art methods.

The ability to localize a robot is an important capability and matching of observations under substantial changes is a prerequisite for robust long-term operation. This paper investigates the problem of efficiently coping with seasonal changes in image data. We present an extension of a recent approach [15] to visual image matching using sequence information. Our extension allows for exploiting GPS priors in the matching process to overcome the main computational bottleneck of the previous method and to handle loops within the image sequences. We present an experimental evaluation using real world data containing substantial seasonal changes and show that our approach outperforms the previous method in case a noisy GPS pose prior is available.

In this paper, we present an appearance-based visual SLAM approach that focuses on detecting loop closures across seasons. Given two image sequences, our method first extracts one descriptor per image for both sequences using a deep convolutional neural network. Then, we compute a similarity matrix by comparing each image of a query sequence with a database. Finally, based on the similarity matrix, we formulate a flow network problem and compute matching hypotheses between sequences. In this way, our approach can handle partially matching routes, loops in the trajectory and different speeds of the robot. With a matching hypothesis as loop closure information and the odometry information of the robot, we formulate a graph based SLAM problem and compute a joint maximum likelihood trajectory.

Image-based localization is an important problem in robotics and an integral part of visual mapping and navigation systems. An approach to robustly match images to previously recorded ones must be able to cope with seasonal changes especially when it is supposed to work reliably over long periods of time. In this paper, we present a novel approach to visual localization of mobile robots in outdoor environments, which is able to deal with substantial seasonal changes. We formulate image matching as a minimum cost flow problem in a data association graph to effectively exploit sequence information. This allows us to deal with non-matching image sequences that result from temporal occlusions or from visiting new places. We present extensive experimental evaluations under substantial seasonal changes. Our approach achieves accurate matching across seasons and outperforms existing state-of-the-art methods such as FABMAP2 and SeqSLAM.

In this paper, we present an approach that allows a quadrocopter to follow a person and to recognize simple gestures using an onboard depth camera. This enables novel applications such as hands-free filming and picture taking. The problem of tracking a person with an onboard camera however is highly challenging due to the self-motion of the platform. To overcome this problem, we stabilize the depth image by warping it to a virtual-static camera, using the estimated pose of the quadrocopter obtained from vision and inertial sensors using an Extended Kalman filter. We show that such a stabilized depth video is well suited to use with existing person trackers such as the OpenNI tracker. Using this approach, the quadrocopter not only obtains the position and orientation of the tracked person, but also the full body pose - which can then for example be used to recognize hand gestures to control the quadrocopter's behaviour. We implemented a small set of example commands (“follow me”, “take picture”, “land”), and generate corresponding motion commands. We demonstrate the practical performance of our approach in an extensive set of experiments with a quadrocopter. Although our current system is limited to indoor environments and small motions due to the restrictions of the used depth sensor, it indicates that there is large potential for such applications in the near future.