Event cameras are novel bio-inspired sensors, which asynchronously capture pixel-level intensity changes in the form of "events". Due to their sensing mechanism, event cameras have little to no motion blur, a very high temporal resolution and require significantly less power and memory than traditional frame-based cameras. These characteristics make them a perfect fit to several real-world applications such as egocentric action recognition on wearable devices, where fast camera motion and limited power challenge traditional vision sensors. However, the ever-growing field of event-based vision has, to date, overlooked the potential of event cameras in such applications. In this paper, we show that event data is a very valuable modality for egocentric action recognition. To do so, we introduce N-EPIC-Kitchens, the first event-based camera extension of the large-scale EPIC-Kitchens dataset. In this context, we propose two strategies: (i) directly processing event-camera data with traditional video-processing architectures (E$^2$(GO)) and (ii) using event-data to distill optical flow information (E$^2$(GO)MO). On our proposed benchmark, we show that event data provides a comparable performance to RGB and optical flow, yet without any additional flow computation at deploy time, and an improved performance of up to 4% with respect to RGB only information.
In machine learning, differential privacy and federated learning concepts are gaining more and more importance in an increasingly interconnected world. While the former refers to the sharing of private data characterized by strict security rules to protect individual privacy, the latter refers to distributed learning techniques in which a central server exchanges information with different clients for machine learning purposes. In recent years, many studies have shown the possibility of bypassing the privacy shields of these systems and exploiting the vulnerabilities of machine learning models, making them leak the information with which they have been trained. In this work, we present the 3DGL framework, an alternative to the current federated learning paradigms. Its goal is to share generative models with high levels of $\varepsilon$-differential privacy. In addition, we propose DDP-$\beta$VAE, a deep generative model capable of generating synthetic data with high levels of utility and safety for the individual. We evaluate the 3DGL framework based on DDP-$\beta$VAE, showing how the overall system is resilient to the principal attacks in federated learning and improves the performance of distributed learning algorithms.
One of the main components of an autonomous vehicle is the obstacle detection pipeline. Most prototypes, both from research and industry, rely on lidars for this task. Pointcloud information from lidar is usually combined with data from cameras and radars, but the backbone of the architecture is mainly based on 3D bounding boxes computed from lidar data. To retrieve an accurate representation, sensors with many planes, e.g., greater than 32 planes, are usually employed. The returned pointcloud is indeed dense and well defined, but high-resolution sensors are still expensive and often require powerful GPUs to be processed. Lidars with fewer planes are cheaper, but the returned data are not dense enough to be processed with state of the art deep learning approaches to retrieve 3D bounding boxes. In this paper, we propose two solutions based on occupancy grid and geometric refinement to retrieve a list of 3D bounding boxes employing lidar with a low number of planes (i.e., 16 and 8 planes). Our solutions have been validated on a custom acquired dataset with accurate ground truth to prove its feasibility and accuracy.
Real-time six degree-of-freedom pose estimation with ground vehicles represents a relevant and well studied topic in robotics, due to its many applications, such as autonomous driving and 3D mapping. Although some systems exist already, they are either not accurate or they struggle in real-time setting. In this paper, we propose a fast, accurate and modular LiDAR SLAM system for both batch and online estimation. We first apply downsampling and outlier removal, to filter out noise and reduce the size of the input point clouds. Filtered clouds are then used for pose tracking and floor detection, to ground-optimize the estimated trajectory. The availability of a pre-tracker, working in parallel with the filtering process, allows to obtain pre-computed odometries, to be used as aids when performing tracking. Efficient loop closure and pose optimization, achieved through a g2o pose graph, are the last steps of the proposed SLAM pipeline. We compare the performance of our system with state-of-the-art point cloud based methods, LOAM, LeGO-LOAM, A-LOAM, LeGO-LOAM-BOR and HDL, and show that the proposed system achieves equal or better accuracy and can easily handle even cases without loops. The comparison is done evaluating the estimated trajectory displacement using the KITTI and RADIATE datasets.
In the emerging high mobility Vehicle-to-Everything (V2X) communications using millimeter Wave (mmWave) and sub-THz, Multiple-Input Multiple-Output (MIMO) channel estimation is an extremely challenging task. At mmWaves/sub-THz frequencies, MIMO channels exhibit few leading paths in the space-time domain (i.e., directions or arrival/departure and delays). Algebraic Low-rank (LR) channel estimation exploits space-time channel sparsity through the computation of position-dependent MIMO channel eigenmodes leveraging recurrent training vehicle passages in the coverage cell. LR requires vehicles' geographical positions and tens to hundreds of training vehicles' passages for each position, leading to significant complexity and control signalling overhead. Here we design a DL-based LR channel estimation method to infer MIMO channel eigenmodes in V2X urban settings, starting from a single LS channel estimate and without needing vehicle's position information. Numerical results show that the proposed method attains comparable Mean Squared Error (MSE) performance as the position-based LR. Moreover, we show that the proposed model can be trained on a reference scenario and be effectively transferred to urban contexts with different space-time channel features, providing comparable MSE performance without an explicit transfer learning procedure. This result eases the deployment in arbitrary dense urban scenarios.
Today, Multi-View Stereo techniques are able to reconstruct robust and detailed 3D models, especially when starting from high-resolution images. However, there are cases in which the resolution of input images is relatively low, for instance, when dealing with old photos, or when hardware constrains the amount of data that can be acquired. In this paper, we investigate if, how, and how much increasing the resolution of such input images through Super-Resolution techniques reflects in quality improvements of the reconstructed 3D models, despite the artifacts that sometimes this may generate. We show that applying a Super-Resolution step before recovering the depth maps in most cases leads to a better 3D model both in the case of PatchMatch-based and deep-learning-based algorithms. The use of Super-Resolution improves especially the completeness of reconstructed models and turns out to be particularly effective in the case of textured scenes.
MIMO systems in the context of 6G Vehicle-to-Everything (V2X) will require an accurate channel knowledge to enable efficient communication. Standard channel estimation techniques, such as Unconstrained Maximum Likelihood (U-ML), are extremely noisy in massive MIMO settings, while structured approaches, e.g., compressed sensing, are suited to low-mobility scenarios and are sensitive to hardware impairments. We propose a novel Multi-Vehicular algebraic channel estimation method for 6G V2X based on unsupervised learning which exploits recurrent vehicle passages in typical urban settings. Multiple training sequences are clustered via K-medoids algorithm based on their \textit{algebraic similarity} to retrieve the MIMO channel eigenmodes, which can be used to improve the channel estimates. Numerical results show remarkable benefits of the proposed method in terms of Mean Squared Error (MSE) compared to standard U-ML solution (15 dB less).
Self-driving technology is expected to revolutionize different sectors and is seen as the natural evolution of road vehicles. In the last years, real-world validation of designed and virtually tested solutions is growing in importance since simulated environments will never fully replicate all the aspects that can affect results in the real world. To this end, this paper presents our prototype platform for experimental research on connected and autonomous driving projects. In detail, the paper presents the overall architecture of the vehicle focusing both on mechanical aspects related to remote actuation and sensors set-up and software aspects by means of a comprehensive description of the main algorithms required for autonomous driving as ego-localization, environment perception, motion planning, and actuation. Finally, experimental tests conducted in an urban-like environment are reported to validate and assess the performances of the overall system.
Recently, the trend of incorporating differentiable algorithms into deep learning architectures arose in machine learning research, as the fusion of neural layers and algorithmic layers has been beneficial for handling combinatorial data, such as shortest paths on graphs. Recent works related to data-driven planning aim at learning either cost functions or heuristic functions, but not both. We propose Neural Weighted A*, a differentiable anytime planner able to produce improved representations of planar maps as graph costs and heuristics. Training occurs end-to-end on raw images with direct supervision on planning examples, thanks to a differentiable A* solver integrated into the architecture. More importantly, the user can trade off planning accuracy for efficiency at run-time, using a single, real-valued parameter. The solution suboptimality is constrained within a linear bound equal to the optimal path cost multiplied by the tradeoff parameter. We experimentally show the validity of our claims by testing Neural Weighted A* against several baselines, introducing a novel, tile-based navigation dataset. We outperform similar architectures in planning accuracy and efficiency.
Event cameras are novel bio-inspired sensors, which asynchronously capture pixel-level intensity changes in the form of "events". The innovative way they acquire data presents several advantages over standard devices, especially in poor lighting and high-speed motion conditions. However, the novelty of these sensors results in the lack of a large amount of training data capable of fully unlocking their potential. The most common approach implemented by researchers to address this issue is to leverage simulated event data. Yet, this approach comes with an open research question: how well simulated data generalize to real data? To answer this, we propose to exploit, in the event-based context, recent Domain Adaptation (DA) advances in traditional computer vision, showing that DA techniques applied to event data help reduce the sim-to-real gap. To this purpose, we propose a novel architecture, which we call Multi-View DA4E (MV-DA4E), that better exploits the peculiarities of frame-based event representations while also promoting domain invariant characteristics in features. Through extensive experiments, we prove the effectiveness of DA methods and MV-DA4E on N-Caltech101. Moreover, we validate their soundness in a real-world scenario through a cross-domain analysis on the popular RGB-D Object Dataset (ROD), which we extended to the event modality (RGB-E).