Registration Number
0222U004568,
0121U109526
, R & D reports
Title
Real-time object localization, classification and tracking system using artificial intelligence algorithms
popup.stage_title
Head
Barkalov Olexandr O., Доктор технічних наук
Registration Date
12-09-2022
Organization
Vasyl' Stus Donetsk National University
popup.description2
With the use of 3 datasets, a study of the accuracy of the final system was conducted. Kinetics400, Mimetics and NTU-RGBD datasets with video content and annotation of human poses were used. As a study was analyzed as separate components of the system and the accuracy of the final model. Also, the most probable pose and the top 5 probable poses were checked for accuracy for a broader understanding of the accuracy of the models.
As a result, the context stream achieved high accuracy on the Kinetics and NTU-RGBD datasets, and comparatively low accuracy on the Mimetics dataset. The low accuracy is due to the fact that the models were trained on the Kinetics and NTU-RGBD datasets, accordingly they cope better with camera perspectives and incomplete poses in these datasets, compared to the completely new video format. The pose analysis stream showed low accuracy on the Kinetics and Mimetics datasets but showed high accuracy when working with the NTU-RGBD dataset. The final model showed that the combination of pose streams and video context allows taking the strengths of both approaches, thus obtaining the comparatively highest accuracy of video prediction results from the datasets.
The novelty of the solution is in the fact that the previous methods of integration use a static combination of two data or successive enrichment of information about the pose with the help of information about the external context, which leads to a strong shift in the prediction results towards the context. In contrast, the proposed integration of features can filter the context information and thus can perform reliable action recognition on both full and incomplete context videos.
Product Description
popup.authors
Baiev Artem V.
Boldyrіeva Valeriia Olegivna
Vietrov Oleh S.
Vasylenko Valeriia Yu.
Zelinska Oksana V.
Neskorodeva Tatyana V
Petranova Maryna Yu.
Prokofiev Mykhailo I.
Rymar Pavlo V.
Shtovba Serhii D.
popup.nrat_date
2022-09-12
