Analysis of skeletal kinematics for vision-based motion capturing

Informacje na temat projektu

SeerPredict

Identyfikator umowy o grant: 957112

DOI

10.3030/957112

Projekt został zamknięty

Data podpisania przez KE 15 Września 2020

Data rozpoczęcia 1 Października 2020

Data zakończenia 31 Marca 2022

Finansowanie w ramach

• INDUSTRIAL LEADERSHIP - Innovation In SMEs

Koszt całkowity

Brak danych

Wkład UE

€ 171 187,50

Koordynowany przez

SEERVISION AG
Switzerland

Periodic Reporting for period 1 - SeerPredict (Analysis of skeletal kinematics for vision-based motion capturing)

Okres sprawozdawczy: 2020-10-01 do 2022-03-31

Podsumowanie kontekstu i ogólnych celów projektu

Seervision builds solutions to automate video production workflows, enabling corporations and universities to create more and more high-quality live video content with minimal overhead. Our AI-driven technology detects people in the frame and creates a skeleton based on which it autonomously moves cameras to best frame and follow the on-screen talent in a variety of productions from streamed townhalls to lecture recordings. Just like camera men, we try to predict how a person will move to be able to act quickly enough and keep them framed nicely throughout the production. Interactions between on-screen talent are a key challenge for this, especially on smaller stages, because the machine learning systems can get confused and fail to smoothly keep track of the person that should be filmed.
With SeerPredict we develop novel machine learning algorithms to improve the reliability and quality of the skeleton estimation and prediction resulting in substantial improvements to the talent tracking performance and overall autonomy of Seervision’s video production solution, making high-quality live video more accessible and affordable.

Prace wykonane od początku projektu do końca okresu sprawozdawczego oraz najważniejsze dotychczasowe rezultaty

As part of the SeerPredict project, we have developed, implemented and validated two key improvements of Seervision’s visual perception and tracking solution:
1) We have developed a novel software module that identifies person-to-person occlusions, especially during crossings, as they occur and mitigates their effect on the quality of the extracted skeleton, enhances the robustness and smoothness of the tracking.
2) We have re-designed the architecture of Seervision’s visual perception and tracking solution, incorporating recent advances in the computer vision research literature. This has reduced the delay that needs to be compensated with skeletal prediction, yielding state-of-the-art performance. In addition, it has also reduced the overall computational overhead.
The developed software has been extensively tested under laboratory conditions, during internal productions and demos at Seervision’s in-house studio and finally with selected customers who have seen a clear improvement in the quality and performance of the talent tracking.

Innowacyjność oraz oczekiwany potencjalny wpływ (w tym dotychczasowe znaczenie społeczno-gospodarcze i szersze implikacje społeczne projektu)

The algorithms developed during this project advance the state-of-the-art of subject tracking for video production use cases. Moreover, the developed software module and improved architecture has been rolled-out into production and helps Seervision maintain its position as the state-of-the-art, premium solution. Moreover, the advances during this project enable further initiatives that will have substantial impact on Seervision’s solution and business.

Person-to-person occulsion during testing at Seervision's offices