Abstract: Depth perception is a key aspect of human vision. It is a routine and essential visual task that the human do effortlessly in many daily activities. This has often been associated with stereo vision, but humans have an amazing ability to perceive depth relations even from a single image by using several monocular cues.
In the computer vision field, if image depth information were available, many tasks could be posed from a different perspective for the sake of higher performance and robustness. Nevertheless, given a single image, this possibility is usually discarded, since obtaining depth information has frequently been performed by three-dimensional reconstruction techniques, requiring two or more images of the same scene taken from different viewpoints. Recently, some proposals have shown the feasibility of computing depth information from single images. In essence, the idea is to take advantage of a priori knowledge of the acquisition conditions and the observed scene to estimate depth from monocular pictorial cues. These approaches try to precisely estimate the scene depth maps by employing computationally demanding techniques. However, to assist many computer vision algorithms, it is not really necessary computing a costly and detailed depth map of the image. Indeed, just a rough depth description can be very valuable in many problems.
In this thesis, we have demonstrated how coarse depth information can be integrated in different tasks following alternative strategies to obtain more precise and robust results. In that sense, we have proposed a simple, but reliable enough technique, whereby image scene regions are categorized into discrete depth ranges to build a coarse depth map. Based on this representation, we have explored the potential usefulness of our method in three application domains from novel viewpoints: camera rotation parameters estimation, background estimation and pedestrian candidate generation. In the first case, we have computed camera rotation mounted in a moving vehicle applying two novels methods based on distant elements in the image, where the translation component of the image flow vectors is negligible. In background estimation, we have proposed a novel method to reconstruct the background by penalizing close regions in a cost function, which integrates color, motion, and depth terms. Finally, we have benefited of geometric and depth information available on single images for pedestrian candidate generation to significantly reduce the number of generated windows to be further processed by a pedestrian classifier. In all cases, results have shown that our approaches contribute to better performances.

Abstract: Video alignment consists of integrating multiple video sequences recorded independently into a single video sequence. This means to register both in time (synchronize
frames) and space (image registration) so that the two videos sequences can be fused
or compared pixel–wise. In spite of being relatively unknown, many applications today may benefit from the availability of robust and efficient video alignment methods.
For instance, video surveillance requires to integrate video sequences that are recorded
of the same scene at different times in order to detect changes. The problem of aligning videos has been addressed before, but in the relatively simple cases of fixed or rigidly attached cameras and simultaneous acquisition. In addition, most works rely
on restrictive assumptions which reduce its difficulty such as linear time correspondence or the knowledge of the complete trajectories of corresponding scene points on the images; to some extent, these assumptions limit the practical applicability of the solutions developed until now. In this thesis, we focus on the challenging problem of aligning sequences recorded at different times from independent moving cameras following similar but not coincident trajectories. More precisely, this thesis covers four studies that advance the state-of-the-art in video alignment. First, we focus on analyzing and developing a probabilistic framework for video alignment, that is, a principled way to integrate multiple observations and prior information. In this way, two different approaches are presented to exploit the combination of several purely visual features (image–intensities, visual words and dense motion field descriptor), and
global positioning system (GPS) information. Second, we focus on reformulating the
problem into a single alignment framework since previous works on video alignment
adopt a divide–and–conquer strategy, i.e., first solve the synchronization, and then
register corresponding frames. This also generalizes the ’classic’ case of fixed geometric transform and linear time mapping. Third, we focus on exploiting directly the
time domain of the video sequences in order to avoid exhaustive cross–frame search.
This provides relevant information used for learning the temporal mapping between
pairs of video sequences. Finally, we focus on adapting these methods to the on–line
setting for road detection and vehicle geolocation. The qualitative and quantitative
results presented in this thesis on a variety of real–world pairs of video sequences show that the proposed method is: robust to varying imaging conditions, different image
content (e.g., incoming and outgoing vehicles), variations on camera velocity, and
different scenarios (indoor and outdoor) going beyond the state–of–the–art. Moreover, the on–line video alignment has been successfully applied for road detection and
vehicle geolocation achieving promising results.

Abstract: Anticipating the intentions of vulnerable road users (VRUs) such as pedestrians
and cyclists can be critical for performing safe and comfortable driving maneuvers. This is the case for human driving and, therefore, should be taken into account by systems providing any level of driving assistance, i.e. from advanced driver assistant systems (ADAS) to fully autonomous vehicles (AVs). In this PhD work, we show how the latest advances on monocular vision-based human pose estimation, i.e. those relying on deep Convolutional Neural Networks (CNNs), enable to recognize the intentions of such VRUs. In the case of cyclists, we assume that they follow the established traffic codes to indicate future left/right turns and stop maneuvers with arm signals. In the case of pedestrians, no indications can be assumed a priori. Instead, we hypothesize that the walking pattern of a pedestrian can allow us to determine if he/she has the intention of crossing the road in the path of the egovehicle, so that the ego-vehicle must maneuver accordingly (e.g. slowing down or stopping). In this PhD work, we show how the same methodology can be used for recognizing pedestrians and cyclists’ intentions. For pedestrians, we perform experiments on the publicly available Daimler and JAAD datasets. For cyclists, we did not found an analogous dataset, therefore, we created our own one by acquiring
and annotating corresponding video-sequences which we aim to share with the
research community. Overall, the proposed pipeline provides new state-of-the-art results on the intention recognition of VRUs.

Abstract: The curse of data labeling is a costly bottleneck in supervised deep learning, where large amounts of labeled data are needed to train intelligent systems. In onboard perception for autonomous driving, this cost corresponds to the labeling of raw data from sensors such as cameras, LiDARs, RADARs, etc. Therefore, synthetic data with automatically generated ground truth (labels) has aroused as a reliable alternative for training onboard perception models.
However, synthetic data commonly suffers from synth-to-real domain shift, i.e., models trained on the synthetic domain do not show their achievable accuracy when performing in the real world. This shift needs to be addressed by techniques falling in the realm of domain adaptation (DA).
The semi-supervised learning (SSL) paradigm can be followed to address DA. In this case, a model is trained using source data with labels (here synthetic) and leverages minimal knowledge from target data (here the real world) to generate pseudo-labels. These pseudo-labels help the training process to reduce the gap between the source and the target domains. In general, we can assume accessing both, pseudo-labels and a few amounts of human-provided labels for the target-domain data. However, the most interesting and challenging setting consists in assuming that we do not have human-provided labels at all. This setting is known as unsupervised domain adaptation (UDA). This PhD focuses on applying SSL to the UDA setting, for onboard visual tasks related to autonomous driving. We start by addressing the synth-to-real UDA problem on onboard vision-based object detection (pedestrians and cars), a critical task for autonomous driving and driving assistance. In particular, we propose to apply an SSL technique known as co-training, which we adapt to work with deep models that process a multi-modal input. The multi-modality consists of the visual appearance of the images (RGB) and their monocular depth estimation. The synthetic data we use as the source domain contains both, object bounding boxes and depth information. This prior knowledge is the
starting point for the co-training technique, which iteratively labels unlabeled real-world data and uses such pseudolabels (here bounding boxes with an assigned object class) to progressively improve the labeling results. Along this
process, two models collaborate to automatically label the images, in a way that one model compensates for the errors of the other, so avoiding error drift. While this automatic labeling process is done offline, the resulting pseudolabels can be used to train object detection models that must perform in real-time onboard a vehicle. We show that multi-modal co-training improves the labeling results compared to single-modal co-training, remaining competitive compared to human labeling.
Given the success of co-training in the context of object detection, we have also adapted this technique to a more crucial and challenging visual task, namely, onboard semantic segmentation. In fact, providing labels for a single image
can take from 30 to 90 minutes for a human labeler, depending on the content of the image. Thus, developing automatic labeling techniques for this visual task is of great interest to the automotive industry. In particular, the new co-training framework addresses synth-to-real UDA by an initial stage of self-training. Intermediate models arising from this stage are used to start the co-training procedure, for which we have elaborated an accurate collaboration policy between the two models performing the automatic labeling. Moreover, our co-training seamlessly leverages datasets from different synthetic domains. In addition, the co-training procedure is agnostic to the loss function used to train the semantic segmentation models which perform the automatic labeling. We achieve state-of-the-art results on publicly available benchmark datasets, again, remaining competitive compared to human labeling.
Finally, on the ground of our previous experience, we have designed and implemented a new SSL technique for UDA in the context of visual semantic segmentation. In this case, we mimic the labeling methodology followed by human labelers. In particular, rather than labeling full images at a time, categories of semantic classes are defined and only those are labeled in a labeling pass. In fact, different human labelers can become specialists in labeling different categories. Afterward, these per-category-labeled layers are combined to provide fully labeled images. Our technique is inspired by this methodology since we perform synth-to-real UDA per category, using the self-training stage previously developed as part of our co-training framework. The pseudo-labels obtained for each category are finally
fused to obtain fully automatically labeled images. In this context, we have also contributed to the development of a new photo-realistic synthetic dataset based on path-tracing rendering. Our new SSL technique seamlessly leverages publicly available synthetic datasets as well as this new one to obtain state-of-the-art results on synth-to-real UDA for semantic segmentation. We show that the new dataset allows us to reach better labeling accuracy than previously existing datasets, at the same time that it complements well them when combined. Moreover, we also show that the new human-inspired SSL technique outperforms co-training.

Abstract: 3D geometric information is essential for on-board perception in autonomous driving and driver assistance. Autonomous vehicles (AVs) are equipped with calibrated sensor suites. As part of these suites, we can find LiDARs, which are expensive active sensors in charge of providing the 3D geometric information. Depending on the operational conditions for the AV, calibrated stereo rigs may be also sufficient for obtaining 3D geometric information, being these rigs less expensive and easier to install than LiDARs. However, ensuring a proper maintenance and calibration of these types of sensors is not trivial. Accordingly, there is an increasing interest on performing monocular depth estimation (MDE) to obtain 3D geometric information on-board. MDE is very appealing since it allows for appearance and depth being on direct pixelwise correspondence without further calibration. Moreover, a set of single cameras with MDE capabilities would still be a cheap solution for on-board perception, relatively easy to integrate and maintain in an AV.
Best MDE models are based on Convolutional Neural Networks (CNNs) trained in a supervised manner, i.e., assuming pixelwise ground truth (GT). Accordingly, the overall goal of this PhD is to study methods for improving CNN-based MDE accuracy under different training settings. More specifically, this PhD addresses different research questions that are described below. When we started to work in this PhD, state-of-theart methods for MDE were already based on CNNs. In fact, a promising line of work consisted in using image-based semantic supervision (i.e., pixel-level class labels) while training CNNs for MDE using LiDAR-based supervision (i.e., depth). It was common practice to assume that the same raw training data are complemented by both types of supervision, i.e., with depth and semantic labels. However, in practice, it was more common to find heterogeneous datasets with either only depth supervision or only semantic supervision. Therefore, our first work was to research if we could train CNNs for MDE by leveraging depth and semantic information from heterogeneous datasets. We show that this is indeed possible, and we surpassed the state-of-the-art results on MDE at the time we did this research. To achieve our results, we proposed a particular CNN architecture and a new training protocol.
After this research, it was clear that the upper-bound setting to train CNN-based MDE models consists in using LiDAR data as supervision. However, it would be cheaper and more scalable if we would be able to train such models from monocular sequences. Obviously, this is far more challenging, but worth to research. Training MDE models using monocular sequences is possible by relying on structure-from-motion (SfM) principles to generate self-supervision. Nevertheless, problems of camouflaged objects, visibility changes, static-camera intervals, textureless areas, and scale ambiguity, diminish the usefulness of such self-supervision. To alleviate these problems, we perform MDE by virtual-world supervision and real-world SfM self-supervision. We call our proposalMonoDEVSNet. We compensate the SfM self-supervision limitations by leveraging
virtual-world images with accurate semantic and depth supervision, as well as addressing the virtual-to-real domain gap. MonoDEVSNet outperformed previous MDE CNNs trained on monocular and even stereo sequences. We have publicly released MonoDEVSNet at <https://github.com/HMRC-AEL/MonoDEVSNet>.
Finally, since MDE is performed to produce 3D information for being used in downstream tasks related to on-board perception. We also address the question of whether the standard metrics for MDE assessment are a good indicator for future MDE-based driving-related perception tasks. By using 3D object detection on point clouds as proxy of on-board perception, we conclude that, indeed, MDE evaluation metrics give rise to a ranking of methods which reflects relatively well the 3D object detection results we may expect.