This paper presents a study of Radio Frequency LoRa device classification performance under challenging channel's variation using a hybrid CNN-LSTM-Attention neural architecture. By addressing the temporal dynamics introduced by Doppler effects in mobile scenarios, combined with targeted data augmentation strategies, our method achieves 99.6% classification accuracy across 10 devices in stationary conditions and maintains robust performance of 85.8% even under high mobility conditions (100 Hz Doppler shift). The proposed hybrid architecture leverages convolutional layers for spatial feature extraction, LSTM layers for modeling temporal dependencies in RF emissions, and an attention mechanism to focus on the most discriminative temporal segments of the signal. Our experimental results, conducted using a dataset of 30 commercial LoRa IoT devices, demonstrate significant performance improvements over state-of-the-art approaches, particularly in challenging mobile environments where Doppler effects typically degrade classification reliability. The model maintains 93.1% accuracy in Line-of-Sight (LOS) mobile scenarios and 87.6% in Non-Line-of-Sight (NLOS) mobile environments, outperforming previous approaches by 3.1% and 2.6% respectively. This work contributes to the field of physical-layer security by demonstrating how temporal modeling techniques can enhance RF fingerprinting performance in realistic mobile deployment scenarios.

Enhancing the security of wireless networks involves implementing a user authentication method when the fingerprint of a network device is unknown or considered a potential threat. This technique is known as radio frequency (RF) fingerprinting. This paper presents a novel method for RF fingerprinting of Internet of Things (IoT) devices, addressing the challenges of the radio frequency spectrum. The proposed architecture integrates a feature generator module that transforms time-series I/Q samples into a multi-dimensional matrix and a deep learning module inspired by the ResNet-50-1D model. We assess the effectiveness of our approach by analyzing a real-world dataset of BT emissions obtained from 10 commercial IoT devices in two challenging indoor environments. The datasets, made publicly accessible on IEEE Dataport, were gathered using a USRP X300 software-defined radio (SDR) in both line-of-sight (LoS) and rich multipath propagation scenarios. Our method showcases excellent results in the TTS scenario and shows promise in the challenging TTD scenario, considering the complex nature of frequency hopping. The evaluation results emphasize the significance of evaluating RF fingerprinting models in various scenarios and offer valuable insights into the strengths and limitations of our approach in handling radio frequency waveforms.

The fusion of multiple images from different modalities is the process of generating a single output image that combines the useful information of all input images. Ideally, the information-rich content of each input image would be preserved, and the cognitive effort required by the user to extract this information should be smaller on the fused image than the one required to examine all images. We propose MobileFuse, an edge computing method targeted at processing large amount of imagery in a bandwidth limited environment using depthwise separable Deep Neural Networks (DNNs). The proposed approach is a hybrid between generative and blending based methods. Our approach can be applied in various fields which require low latency interaction with the user or with an autonomous system. The main challenge in training DNNs for image fusion is the sparsity of data with representative ground truth. Registering images from different sensors is a major challenge in itself, and generating a ground truth from them is another massive one. For this reason, we also propose a multi-focus and multi-lighting framework to generate training dataset using unregistered images. We show that our edge network can perform faster than its state-of-the-art baseline, while improving the fusion quality.

Multimodal image registration is a challenging task. To begin with, the variation of parallax in the images makes the process intrinsically tricky. Additionally, due to phenomenology differences in modalities, the appearance of the same feature may vary significantly between the images making the registration laborious. To help mitigate these issues, we propose a two-step approach targeted at visible and infrared imagery. First, we train a generative adversarial network to learn the domain transfer function between the visible and the infrared domain, thereby mitigating the impact of the visual dissimilarity between the images. Second, we train a deep Siamese network to compute a homography in an unsupervised setting. Both elements are combined and trained sequentially. Our method is evaluated on a publicly available dataset. Our results show that the proposed method provides a reduction of more than 30% on average from the previous state-of-the-art, and outperforms several baselines and recent deep homography methods.

Cognitive vehicles (CV) differ from smart vehicles (SV) in a way that they don't just rely on the sensors' readings and follow rigorously the patterns and functions already preprogrammed externally. CVs utilize the different sensors as a source of information, which needs to be processed and turned into intelligence and perception. CVs learn at a scale, make assumptions, predict outcomes, and learn from experience rather than being explicitly programmed. In this work, we attempt to present a model that duplicates the cognitive process through which humans can self-localize. We present an innovative GNSS-free solution for vehicle self-localization based on detection pattern recognition of visual anchors. The proposed cognitive approach is successfully tested in different routes taken from a real urban environment. The system location estimates are compared with the GPS reported locations and show promising performances.

Augmented reality devices require multiple sensors to perform various tasks such as localization and tracking. Currently, popular cameras are mostly frame-based (e.g. RGB and Depth) which impose a high data bandwidth and power usage. With the necessity for low power and more responsive augmented reality systems, using solely frame-based sensors imposes limits to the various algorithms that needs high frequency data from the environement. As such, event-based sensors have become increasingly popular due to their low power, bandwidth and latency, as well as their very high frequency data acquisition capabilities. In this paper, we propose, for the first time, to use an event-based camera to increase the speed of 3D object tracking in 6 degrees of freedom. This application requires handling very high object speed to convey compelling AR experiences. To this end, we propose a new system which combines a recent RGB-D sensor (Kinect Azure) with an event camera (DAVIS346). We develop a deep learning approach, which combines an existing RGB-D network along with a novel event-based network in a cascade fashion, and demonstrate that our approach significantly improves the robustness of a state-of-the-art frame-based 6-DOF object tracker using our RGB-D-E pipeline. Our code and our RGB-D-E evaluation dataset are available at https://github.com/lvsn/rgbde-tracking.

We present AWARE, a video monitoring library at the heart of a camera-based Air Traffic Control (ATC) System already deployed in several airports. The system makes use of a network of visible or thermal cameras for detection, tracking and positioning of moving objects on the runway and apron areas. We discuss the main challenges encountered by the system along with the implemented solutions and present the system performances in selected examples of a typical installation.