David Anastasiu
Santa Clara University, CA·Computer Science·Early-career
Cameras now watch streets, warehouses and workplaces almost everywhere, but they are mostly used to review what already went wrong. The question driving this project is whether a system can watch a video stream live and predict that something bad is about to happen in time to prevent it: warning a driver, applying a car's brakes, or halting an escalator.
- Fields
Computer Science
Engineering
- Topics
video anomaly detection
anomaly anticipation
explainable AI
- Grants
- 1 known grant, through May 2030
What they work on
- Create a digital-twin framework that semi-automatically captures and annotates video anomaly anticipation datasets from static camera viewpoints, recomposing real objects and motion patterns into new motion anomalies.
- Use generative AI models conditioned on designed anomalous motion patterns to synthesize realistic videos containing anomalies, supplying training data that real-world footage cannot provide at scale.
- Design real-time multimodal models that forecast object trajectories and positions and from them estimate the probability of an upcoming motion anomaly before it occurs.
- Develop descriptive visual AI models that produce understandable visual or text explanations of anticipated anomalies for the stakeholders who must respond.
- Validate the full pipeline on traffic anomaly anticipation as a case study, testing whether predictions arrive early enough and accurately enough to trigger preventive action.
More about the research
Just as important, the system should be able to say why it acted, in pictures or words, so the people affected can judge whether the intervention made sense. Anticipating anomalies, as opposed to detecting them after the fact, is largely unexplored in video AI, and three obstacles stand in the way: there are almost no real-world datasets annotated for anticipation, there are no fast enough algorithms for prediction in real time, and AI decisions in this setting are hard to explain.
Why it matters
Shifting video AI from after-the-fact detection to advance warning could prevent harm rather than document it: fewer pedestrian injuries and deaths in intelligent traffic systems, fewer equipment accidents and losses in warehouses, and an added safeguard in settings such as air traffic control. Explanations accompanying each alert would let operators and affected people understand and contest automated interventions instead of simply being subject to them.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Hamed Tabkhi
University of North Carolina at Charlotte, NC·Computer Science
Develop a semi-supervised transformer model that detects unusual or threatening behavior in real time from surveillance video. Use identity-neutral visual features so that no personally identifiable information is needed.
Xuegang Ban
University of Washington, WA·Engineering
Adapt and scale privacy-preserving techniques so that traffic, trajectory, and naturalistic driving data can be shared securely under both centralized and distributed data-sharing models, and evaluate them on real-world transportation datasets.
Mitsunori Ogihara
University of Miami, FL·Computer Science
Develop accelerated methods for analyzing medical images, using AI/ML together with coupled software and programmable hardware so results arrive fast enough for real-time use.
The school
Santa Clara University: 8 labs on PI Pages, $8M a year on research, 24 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.