Dataset: baza_rukomet_2023.zip, 1.49 GB Access Condition: Access restricted to higher education institution's students and staff in RH Description: A handball action recognition dataset (English)
Please login to the repository to save this object to your list.
Cite this document
Host, K., Ivašić Kos, M. & Pobar, M. (2023). UNIRI-HBD_v2 Dataset [Data set]. https://urn.nsk.hr/urn:nbn:hr:195:291879.
Host, Kristina, et al. UNIRI-HBD_v2 Dataset. Fakultet informatike i digitalnih tehnologija, 2023. 27 Nov 2024. https://urn.nsk.hr/urn:nbn:hr:195:291879.
Host, Kristina, Marina Ivašić Kos, and Miran Pobar. 2023. UNIRI-HBD_v2 Dataset. Fakultet informatike i digitalnih tehnologija. https://urn.nsk.hr/urn:nbn:hr:195:291879.
Host, K., Ivašić Kos, M. and Pobar, M. 2023. UNIRI-HBD_v2 Dataset. Fakultet informatike i digitalnih tehnologija. [Online]. [Accessed 27 November 2024]. Available from: https://urn.nsk.hr/urn:nbn:hr:195:291879.
Host K, Ivašić Kos M, Pobar M. UNIRI-HBD_v2 Dataset. [Internet]. Fakultet informatike i digitalnih tehnologija: , HR; 2023, [cited 2024 November 27] Available from: https://urn.nsk.hr/urn:nbn:hr:195:291879.
K. Host, M. Ivašić Kos and M. Pobar, UNIRI-HBD_v2 Dataset, Fakultet informatike i digitalnih tehnologija, 2023. Accessed on: Nov 27, 2024. Available: https://urn.nsk.hr/urn:nbn:hr:195:291879.
Miran Pobar Fakultet informatike i digitalnih tehnologija
Scientific / art field, discipline and subdiscipline
SOCIAL SCIENCES Information and Communication Sciences Information Systems and Information Science
Abstract (english)
High quality video footages were captured both on indoors and outdoors handball courts during various training sessions and competitions of the Croatian handball club Kvarner Kostrena, with actions and activities performed by young male and female players. The recordings were made in full HD resolution (1920x1080) with a frame rate between 30 and 60 FPS (frames per second) using 3 or 4 stationary GoPRO cameras placed at a height of 1.5 m on the left and right edges of the court. From the spectators' point of view, the height of the camera, which was 10 m from the edge of the pitch, was at 3.5 m.
The obtained footages pose a significant challenge because there is a cluttered background caused by multiple players involved in the scene, who often occlude each other and wear jerseys of a similar color to the background, which can make them difficult to recognize Furthermore, the players may be located at different distances from the camera and performing different actions and activities at the same time, which makes the further process of creating the handball database for action recognition quite difficult.
In total there are 5389 video clips from 10 different classes: Catch (562), Throw (556), Background (806), Passing(715), Defense (471), Jump-Shot (520), Shot (351), Crossing (306), Dribbling (525), and Running (577).
The videos from all classes were randomly divided into training and testing sets in a ratio of 80:20.
Methods (english)
To obtain a large amount of ground-truth data for the handball domain, we firstly collected data on the handball sports fields using different GoPro cameras, in the next step we detected the players on the videos using object detection algorithm to obtain location information, that is, bounding boxes around the players, and then tracked the detected players using tracking-by-detection algorithm in order to assign to the bounding boxes a label with a corresponding ID. It is important to point out that players detection and tracking is mandatory to later recognize ongoing actions as the actions are performed by the players themselves and last for a sequence of consecutive frames. The basic handball actions were defined and then temporally annotated using the free video annotation research tool Anvil. The next step of data preparation merges the tracked player with the action taking place and involves manual validation of the obtained results to obtain the final ground-truth data.
Number: IP-2016-06-8345 Title (croatian): Automatsko raspoznavanje akcija i aktivnosti u multimedijalnom sadržaju iz domene sporta Title (english): Automatic recognition of actions and activities in multimedia content from sports domain Acronym: RAASS Leader: Marina Ivašić Kos Jurisdiction: Croatia Funding stream: IP
Project
Number: uniri-drustv-18-222 Title (croatian): Automatsko raspoznavanje sportskih tehnika kod mladih sportaša i rekreativaca u svrhu usvajanja motoričkih vještina i usavršavanje stila Title (english): Automatic recognition of sports technique in young athletes and amateurs for the purpose of adoption of motor skills and style enhancement Leader: Marina Ivašić Kos Jurisdiction: Croatia Funding stream: Sveučilište u Rijeci