Vincenzo Carletti

The reliable detection of unconventional obstacles, such as rocks, is essential for ensuring safe and continuous railway operations. However, this task is challenging due to the visual heterogeneity of rocks and the scarcity of annotated railway data. Existing vision-based systems often rely on conventional classifiers, whose generalization ability is limited in this setting. In this work, we introduce Textual-Visual Alignment for Rock Detection (TVA-Rock), a parameter-efficient textual-visual alignment framework that leverages the representational strength of vision-language models for rock detection within a two-stage pipeline comprising track segmentation and patch-wise classification. TVA-Rock employs a sequential textual-visual prompt alignment strategy, in which textual prompts are first optimized to encode task-relevant semantics, and subsequently used to guide visual prompt tuning, yielding rock-aware visual representations without modifying backbone weights. This sequential design stabilizes multimodal optimization, mitigates overfitting in data-scarce regimes, and enables efficient inference by discarding the textual branch after training. Extensive experiments on realistic onboard-camera samples show that TVA-Rock outperforms existing approaches, that are also more complex, achieving state-of-the-art performance in identifying small and visually irregular rocks. Ablation studies further validate the incremental contribution of textual prompting, visual prompting, and architectural generalization across different VLM backbones. These results demonstrate the potential of the proposed method as a robust and efficient strategy for the detection of rocks on railway tracks.

Recent developments in the railway domain have outlined the need to design reliable automatic monitoring systems, whose applicability is strongly dependent on the capability to timely detect potential obstacles on the tracks, including unusual ones such as rocks: even if typical of the railway environment, rocks are still challenging to recognize with traditional object detectors. In this paper, we propose a two-stage approach for the recognition of rocks on railway tracks, in which the region of interest, identified through a semantic segmentation network, is divided into patches analyzed by a CNN-based binary classifier, without sacrificing localization capability. To achieve a performance assessment aligned with real application requirements, the proposed system was tested on both static images and videos collected by a camera mounted on board the train, to which synthetically generated rocks were added, modifying their size and position in order to follow the perspective of the train’s movement. The experimentation conducted, over both still images and videos, confirm the effectiveness of the proposed approach, which confirms to be a viable solutions for real time rocks detection in real scenarios.

Object detection and recognition have undergone a profound transformation over the past decade, driven by the rapid evolution of deep learning architectures, the availability of large annotated datasets, and the increasing computational power of modern hardware

Automatic fire detection in video surveillance is a critical capability for safeguarding industrial sites, public infrastructures, and natural environments. Although recent advances in deep learning have substantially improved early fire detection, the performance of existing methods remains highly dependent on scene conditions such as viewing range, background activity, and the presence of fire-like distractors. This paper presents a comprehensive evaluation of real-time fire detection techniques on the ONFIRE and FIRE-TASTIC benchmarks, currently the most extensive datasets available for video-based fire monitoring. Through a systematic comparison of methods that differ in detector quality, temporal modeling, and semantic reasoning capabilities, we highlight the strengths and limitations of current design paradigms and analyze their behaviour across scenarios that differ in terms of range and background activity. Our results reveal clear performance trends, emphasizing the crucial role of careful detector training, scenario-aware adaptation, temporal analysis, and lightweight vision–language integration. Building on these insights, we outline promising research directions, including the development of more representative datasets, improved semantic confirmation modules, and robust continual learning strategies capable of maintaining reliability in dynamic real-world deployments.

One of the key advantages of Federated Learning (FL) is its ability to collaboratively train a Machine Learning (ML) model while keeping clients' data on-site. However, this can create a false sense of security. Despite not sharing private data increases the overall privacy, prior studies have shown that gradients exchanged during the FL training remain vulnerable to Gradient Inversion Attacks (GIAs). These attacks allow reconstructing the clients' local data, breaking the privacy promise of FL. GIAs can be launched by either a passive or an active server. In the latter case, a malicious server manipulates the global model to facilitate data reconstruction. While effective, earlier attacks falling under this category have been demonstrated to be detectable by clients, limiting their real-world applicability. Recently, novel active GIAs have emerged, claiming to be far stealthier than previous approaches. This work provides the first comprehensive analysis of these claims, investigating four state-of-the-art GIAs. We propose novel lightweight client-side detection techniques, based on statistically improbable weight structures and anomalous loss and gradient dynamics. Extensive evaluation across several configurations demonstrates that our methods enable clients to effectively detect active GIAs without any modifications to the FL training protocol.