Abstract
Understanding and monitoring wildlife populations is critical for conservation,
yet manual observation is labor-intensive and limited in scope. This
study evaluates the effectiveness of modern tracking-by-detection algorithms
for monitoring red deer (Cervus elaphus). Using the MammAlps dataset,
YOLOv8 was trained as the detection backbone, achieving excellent results
with precision and recall above 99% and mAP@0.5-0.95 of 96.1%. The detected
outputs were then integrated with four trackers: Centroid, BoT-SORT,
ByteTrack, and StrongSORT, and evaluated using Multiple Object Tracking
(MOT) metrics. The results showed that BoT-SORT performed best (MOTA =
98.6%), followed by ByteTrack and StrongSORT, although all trackers achieved
low IDF1 scores, emphasizing persistent identity fragmentation challenges.
The findings align with prior ecological benchmarks and confirm that while
detection is highly reliable, identity preservation remains limited. This research
establishes a baseline performance for the tracking of red deer and
highlights the need for adapted models to advance conservation monitoring.