Abstract
Whilst species detection and individual identification of great apes have previously been attempted with success, research into automated behaviour recognition remains a humancentric task. In this paper we present a first great ape-specific visual behaviour recognition system and related data annotations for detecting core ape behaviours. Among others, these include sitting, walking, and climbing. The presented basic dual-stream model with late fusion is capable of performing multi-subject multi-behaviour recognition on apes in challenging camera trap footage. More than 180,000 frames across 500 videos from the PanAfrican dataset were annotated with individual IDs and
behaviour labels to end-to-end train and evaluate the system. In summary, our key contributions include a proposed system capable of an accuracy of 73.52%, along with the behaviour annotated dataset, key code and network weights.
behaviour labels to end-to-end train and evaluate the system. In summary, our key contributions include a proposed system capable of an accuracy of 73.52%, along with the behaviour annotated dataset, key code and network weights.
| Original language | English |
|---|---|
| Publication status | Published - 15 Jan 2021 |
| Event | International Conference on Pattern Recognition (ICPR) Workshop on Visual Observation and Analysis of Vertebrate And Insect Behavior - Milan, Italy Duration: 10 Jan 2021 → 15 Jan 2021 Conference number: 25 https://www.micc.unifi.it/icpr2020/ |
Conference
| Conference | International Conference on Pattern Recognition (ICPR) Workshop on Visual Observation and Analysis of Vertebrate And Insect Behavior |
|---|---|
| Abbreviated title | VAIB |
| Country/Territory | Italy |
| City | Milan |
| Period | 10/01/21 → 15/01/21 |
| Internet address |