Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Pose estimation for the critically endangered African penguin
Thesis (MSc)--Stellenbosch University, 2025.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Stellenbosch : Stellenbosch University
2026
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867614026055286785 |
|---|---|
| access_status_str | Open Access |
| author | Van den Berg, Matthew Shane |
| author2 | Dufourq, Emmanuel |
| author_browse | Dufourq, Emmanuel Van den Berg, Matthew Shane |
| author_facet | Dufourq, Emmanuel Van den Berg, Matthew Shane |
| author_sort | Van den Berg, Matthew Shane |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (MSc)--Stellenbosch University, 2025. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/134842 |
| institution | Stellenbosch University (South Africa) |
| last_indexed | 2026-06-10T12:45:29.584Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2026 |
| publishDateRange | 2026 |
| publishDateSort | 2026 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/134842 Pose estimation for the critically endangered African penguin Van den Berg, Matthew Shane Dufourq, Emmanuel Jeantet, Lorène Stellenbosch University. Faculty of Science. Dept. of Applied Mathematics. African penguins -- Behavior Graphics processing units Penguins -- Monitoring Endangered species Thesis (MSc)--Stellenbosch University, 2025. Van den Berg, M. S. 2025. Pose Estimation for the Critically Endangered African Penguin. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/b9d33a69-70af-4669-b328-4c44db8fb0a0 ENGLISH ABSTRACT: Pose estimation tracks predefined anatomical points, or keypoints, in images and video. Applied to wildlife, animal pose estimation (APE) can automate behavioural monitoring, providing valuable insights for both ecological studies and conservation efforts. African penguins (Spheniscus demersus), being endemic to Southern Africa, critically endangered, and of ecological and economic value, were selected as a test species. Deep learning based pose estimation methods typically use an encoder-decoder architecture, where the encoder extracts spatial and semantic features from an input image, and the decoder transforms these into a representation of the animal body. Current stateof- the-art (SOTA) APE systems employ detection-based decoders that are architecturally complex, computationally expensive, and too slow for continuous, real-time monitoring on resource constrained devices. Regression-based decoders are typically architecturally simpler, less computationally expensive, and possess lower inference speeds but are underexplored in APE research. This study develops regression-based APE models for African penguins, prioritising simplicity, small model size, fast inference time, and low computational cost while matching SOTA keypoint localisation accuracy. Two datasets with varying levels of difficulty for APE were collected: one containing a single penguin per frame, named Simple Single Penguin (SSP), and another containing multiple penguins per frame, named Aquarium Multi-penguin (AMP). To improve keypoint localisation accuracy, several lightweight convolutional encoders and a range of loss functions, including novel variants, were tested. The resulting models were compared with two leading SOTA APE methods. The regression-based APE models remained competitive in keypoint localisation accuracy compared to SOTA approaches. The best regression-based models outperformed SOTA on SSP (PCK@0.1 = 0.8843 vs 0.7547) and trailed by 0.0392 on AMP (PCK@0.1 = 0.6827 vs 0.7219). PCK@0.1 gives the percentage of keypoints predicted within 0.1 times the Euclidean distance of the body segment of the animal from the ground truth keypoint. The most efficient regression-based model required over two orders of magnitude fewer giga floating-point operations per second (GFLOPs) than the SOTA models for SSP (0.31 GFLOPs vs 59.68 GFLOPs). For AMP, this difference remained substantial, with GFLOPs more than an order of magnitude lower (0.31 GFLOPs vs 5.83 GFLOPs). Graphics processing unit (GPU) and central processing unit (CPU) inference times for the fastest regression-based models were more than ten times faster than those of the SOTA models for SSP (3.15 ms vs 37.52 ms GPU, and 9.13 ms vs 290.02 ms CPU). For AMP, GPU inference times were more than three times faster and CPU inference times more than 4.5 times faster. Additionally, a novelly developed loss function, named simple-angle loss, improved keypoint localisation on SSP. It yielded an average increase of 0.0464 in PCK@0.1 across all encoder architectures tested compared to the mean absolute error loss, a common loss function for regression-based pose estimation used to establish baseline performance. These results show that regression-based APE can deliver order-of-magnitude gains in speed and efficiency with comparable keypoint localisation accuracy under the conditions presented in this study, warranting further investigation for real-time wildlife monitoring. AFRIKAANSE OPSOMMING: Houdingskatting is n tegniek om voorafgedefinieërde skeletpunte (sogenaamde sleutelpunte) na te speur in beelde en videos. Wanneer houdingskatting gebruik word op beeldmateriaal van diere, kan gedragsmonitering outomaties gedoen word. Dit kan waardevolle insigte verskaf in ekologiese studies, en gevolglik in bewaringspogings. Die Brilpikkewyn spesie (Spheniscus demersus) is endemies aan suidelike Afrika, en is gekies as n toetsspesie vir hierdie werk. Hierdie spesie word geklassifiseer as krities bedreig, en is ook van ekologiese en ekonomiese belang. Diepleer-gebaseerde houdingskattingsmetodes gebruik tipies n enkodeerdekodeer argitektuur. Hier onttrek die enkodeerder sematiese en posisionele kenmerke uit n gegewe toevoerbeeld, sodat die dekodeerder hierdie inligting dan verwerk in n voorstelling van die liggaam van die dier. Die voorste dierehoudingskattingstelsels gebruik sogenaamde opsporingsgebaseerde dekodeerders. Hierdie stelsels benodig komplekse argitekture, vereis oormatige berekeningsbronne, en is te stadig om te gebruik in die geval van aaneenlopende, reëletyd monitering op bronbeperkte tegnologie. Daarteenoor het regressiegebaseerde dekodeerders tipies n eenvoudiger argitektuur, benodig minder berekeningsbronne, en het laer inferensietye. Hierdie dekodeerders is egter minder algemeen in navorsing oor dierehoudingskatterstelsels. Hierdie navorsing ontwikkel dan juis regressiegebaseerde dierehoudingskattingsmodelle spesifiek vir die Brilpikkewyn. Die fokus van die studie beklemtoon stelsels met eenvoud, klein modelgroottes, vinnige inferensietye het en lae berekeningskoste, wat net so akkuraat sal wees as die opsporingsgebaseerde stelsels. Twee datastelle met verskillende kompleksiteitsvlakke is versamel: die eerste (SSP) bevat n enkele pikkewyn op elke videogreep, en die tweede (AMP) bevat verskeie pikkewyne per videogreep. Om die sleutelpuntopsporing se akkuraatheid te verbeter, is verskeie berekeningsligte konvolusionele dekodeerders en n reeks verliesfunksies (insluitende nuwe unieke variasies) getoets. Die resulterende modelle is vergelyk met die twee beste bestaande opsporingsgebaseerde stelsels vir dierehoudingskatting.berekeningskoste, wat net so akkuraat sal wees as die opsporingsgebaseerde stelsels. Twee datastelle met verskillende kompleksiteitsvlakke is versamel: die eerste (SSP) bevat n enkele pikkewyn op elke videogreep, en die tweede (AMP) bevat verskeie pikkewyne per videogreep. Om die sleutelpuntopsporing se akkuraatheid te verbeter, is verskeie berekeningsligte konvolusionele dekodeerders en n reeks verliesfunksies (insluitende nuwe unieke variasies) getoets. Die resulterende modelle is vergelyk met die twee beste bestaande opsporingsgebaseerde stelsels vir dierehoudingskatting. 2026-01-12T10:04:50Z 2026-01-12T10:04:50Z 2025-12 Thesis https://scholar.sun.ac.za/handle/10019.1/134842 Stellenbosch University xviii, 163 pages : illustrations application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | African penguins -- Behavior Graphics processing units Penguins -- Monitoring Endangered species Van den Berg, Matthew Shane Pose estimation for the critically endangered African penguin |
| title | Pose estimation for the critically endangered African penguin |
| title_full | Pose estimation for the critically endangered African penguin |
| title_fullStr | Pose estimation for the critically endangered African penguin |
| title_full_unstemmed | Pose estimation for the critically endangered African penguin |
| title_short | Pose estimation for the critically endangered African penguin |
| title_sort | pose estimation for the critically endangered african penguin |
| topic | African penguins -- Behavior Graphics processing units Penguins -- Monitoring Endangered species |
| url | https://scholar.sun.ac.za/handle/10019.1/134842 |
| work_keys_str_mv | AT vandenbergmatthewshane poseestimationforthecriticallyendangeredafricanpenguin |