A5033120247- Video Surveillance and Tracking Methods
- Human Pose and Action Recognition
- Anomaly Detection Techniques and Applications
- Advanced Neural Network Applications
- Advanced Vision and Imaging
- Advanced Image and Video Retrieval Techniques
- Domain Adaptation and Few-Shot Learning
- Visual Attention and Saliency Detection
- Seismology and Earthquake Studies
- Time Series Analysis and Forecasting
- Image Enhancement Techniques
- 3D Shape Modeling and Analysis
- Human Mobility and Location-Based Analysis
- Impact of Light on Environment and Health
- Seismic Waves and Analysis
- Multimodal Machine Learning Applications
- Remote Sensing and LiDAR Applications
- Autonomous Vehicle Technology and Safety
- Medical Image Segmentation Techniques
- Building Energy and Comfort Optimization
- earthquake and tectonic studies
- Computer Graphics and Visualization Techniques
- Earthquake Detection and Analysis
- Automated Road and Building Extraction
- Network Security and Intrusion Detection
Sapienza University of Rome
2020-2024
University of Padua
2022-2023
Azienda Unità Sanitaria Locale 11 di Empoli
2023
University of Bern
2023
Azienda Ospedaliero-Universitaria Careggi
2023
Agostino Gemelli University Polyclinic
2023
Università Cattolica del Sacro Cuore
2023
Universidade Federal do Rio Grande do Norte
2022
National Research Council
2022
Boston University
2022
Most recent successes on forecasting the people motion are based LSTM models and all most progress has been achieved by modelling social interaction among with scene. We question use of propose novel Transformer Networks for trajectory forecasting. This is a fundamental switch from sequential step-by-step processing LSTMs to only-attention-based memory mechanisms Transformers. In particular, we consider both original Network (TF) larger Bidirectional (BERT), state-of-the-art natural language...
Human pose forecasting is a complex structured-data sequence-modelling task, which has received increasing attention, also due to numerous potential applications. Research mainly addressed the temporal dimension as time series and interaction of human body joints with kinematic tree or by graph. This decoupled two aspects leveraged progress from relevant fields, but it limited understanding structural joint spatio-temporal dynamics pose.Here we propose novel Space-Time-Separable Graph...
Video segmentation research is currently limited by the lack of a benchmark dataset that covers large variety sub problems appearing in video and enough to avoid over fitting. Consequently, there little analysis which generalizes across subtasks, it not yet clear how should leverage information from still-frames, as previously studied image segmentation, alongside specific information, such temporal volume, motion occlusion. In this work we provide an based on annotations dataset, where each...
This paper proposes a probabilistic graphical model for the problem of propagating labels in video sequences, also termed label propagation problem. Given limited amount hand labelled pixels, typically start and end frames chunk video, an EM based algorithm propagates through rest sequence. As result, user obtains pixelwise sequences along with class probabilities at each pixel. Our novel provides essential tool to reduce tedious labelling thus producing copious amounts useable ground truth...
Person search has recently gained attention as the novel task of finding a person, provided cropped sample, from gallery non-cropped images, whereby several other people are also visible. We believe that i. person detection and re-identification should be pursued in joint optimization framework ii. leverage query image extensively (e.g. emphasizing unique patterns). However, so far, no prior art realizes this. introduce query-guided end-to-end network (QEEPS) to address both aspects. most...
Recent approaches on trajectory forecasting use tracklets to predict the future positions of pedestrians exploiting Long Short Term Memory (LSTM) architectures. This paper shows that adding vislets, is, short sequences head pose estimations, allows increase significantly performance. We then propose vislets in a novel framework called MX-LSTM, capturing interplay between and thanks joint unconstrained optimization full covariance matrices during LSTM backpropagation. At same time, MX-LSTM...
The rapid advances of transportation infrastructure have led to a dramatic increase in the demand for smart systems capable monitoring traffic and street safety. Fundamental these applications are community-based evaluation platform benchmark object detection multi-object tracking. To this end, we organize AVSS2017 Challenge on Advanced Traffic Monitoring, conjunction with International Workshop Street Surveillance Safety Security (IWT4S), evaluate state-of-the-art tracking algorithms...
Federated Learning (FL) deals with learning a central model (i.e. the server) in privacy-constrained scenarios, where data are stored on multiple devices clients). The has no direct access to data, but only updates of parameters computed locally by each client. This raises problem, known as statistical heterogeneity, because clients may have different distributions domains). is partly alleviated clustering clients. Clustering reduce heterogeneity identifying domains, it deprives cluster and...
Earthquake forecasting and prediction have long in some cases sordid histories but recent work has rekindled interest based on advances early warning, hazard assessment for induced seismicity successful of laboratory earthquakes. In the lab, frictional stick-slip events provide an analog earthquakes seismic cycle. Labquakes are ideal targets machine learning (ML) because they can be produced sequences under controlled conditions. Recent works show that ML predict several aspects labquakes...
Anomalies are rare and anomaly detection is often therefore framed as One-Class Classification (OCC), i.e. trained solely on normalcy. Leading OCC techniques constrain the latent representations of normal <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> motions to limited volumes detect abnormal anything outside, which accounts satisfactorily for openset'ness anomalies. But normalcy shares same property since humans can perform action in...
Computational and memory costs restrict spectral techniques to rather small graphs, which is a serious limitation especially in video segmentation. In this paper, we propose the use of reduced graph based on superpixels. contrast previous work, reweighted such that resulting segmentation equivalent, under certain assumptions, full graph. We consider equivalence terms normalized cut its clustering relaxation. The proposed method reduces runtime consumption yields par results image Further, it...
Video segmentation has become an important and active research area with a large diversity of proposed approaches. Graph-based methods, enabling top-performance on recent benchmarks, consist three essential components: 1. powerful features account for object appearance motion similarities; 2. spatio-temporal neighborhoods pixels or superpixels (the graph edges) are modeled using combination those features; 3. video is formulated as partitioning problem. While wide variety have been explored...
In this article, we explore the correlation between people trajectories and their head orientations. We argue that trajectory pose forecasting can be modelled as a joint problem. Recent approaches on leverage short-term (aka tracklets) of pedestrians to predict future paths. addition, sociological cues, such expected destination or pedestrian interaction, are often combined with tracklets. propose MiXing-LSTM (MX-LSTM) capture interplay positions orientations (vislets) thanks unconstrained...
3D object detectors based only on LiDAR point clouds hold the state-of-the-art modern street-view benchmarks. However, LiDAR-based poorly generalize across domains due to domain shift. In case of LiDAR, in fact, shift is not changes environment and appearances, as for visual data from RGB cameras, but also related geometry (e.g., density variations). This paper proposes SF-UDA <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3D</sup> , first...
Transformer Networks have established themselves as the de-facto state-of-the-art for trajectory forecasting but there is currently no systematic study on their capability to model motion patterns of people, without interactions with other individuals nor social context. There abundant literature LSTMs, CNNs and GANs this subject. However methods adopting techniques achieve great performances by complex models a clear analysis adoption plain sequence missing. This paper proposes first...
Detecting the anomaly of human behavior is paramount to timely recognizing endangering situations, such as street fights or elderly falls. However, detection complex since anomalous events are rare and because it an open set recognition task, i.e., what at inference has not been observed training. We propose COSKAD, a novel model that encodes skeletal motion by graph convolutional network learns COntract SKeletal kinematic embeddings onto latent hypersphere minimum volume for Video Anomaly...
We introduce the novel task of Crowd Volume Estimation (CVE), defined as process estimating collective body volume crowds using only RGB images. Besides event management and public safety, CVE can be instrumental in approximating weight, unlocking weight sensitive applications such infrastructure stress assessment, assuring even balance. propose first benchmark for CVE, comprising ANTHROPOS-V, a synthetic photorealistic video dataset featuring diverse urban environments. Its annotations...
State Space Models (SSMs) have recently enjoyed a rise to prominence in the field of deep learning for sequence modeling, especially as an alternative Transformers. Their success stems from avoiding two well-known drawbacks attention-based models: quadratic complexity with respect length and inability model long-range dependencies. The SSM variant Mamba has demonstrated performance comparable Transformers without any form attention, thanks use selective mechanism state parameters....
Fault zone properties evolve dynamically during the seismic cycle due to stress changes, microcracking, and wall rock damage. Understanding these changes is vital gaining insights into earthquake preparation post-seismic processes. The latter include fault healing, which refers recovery of mechanical elastic in zones after aseismic slip.&#160;Despite its importance, detecting characterizing healing through signals remains a challenge subtle nature changes.In this study, we investigate...
Estimation of earthquake parameters has always been a focus for seismologists. Efficient and rapid determination location magnitude is essential mitigating the potential hazards associated with seismic shaking. Nowadays, Earthquake Early Warning Systems (EEWS) are implemented in most earthquake-prone areas, system varying according to specific needs. Although methods their estimation exist, many still lack fast enough process, which crucial reducing waiting time before issuing warning.Here,...