Uma metodologia para detecção e rastreamento de múltiplos cães em baias utilizando YOLOv11 com avaliação de algoritmos de rastreamento

Abstract

Multiple Animal Tracking (MAT), a subfield of Multiple Object Tracking (MOT), is an essential computer vision tool for applications such as behavioral monitoring and animal welfare. However, tracking multiple dogs in kennels using a top-down view presents significant challenges due to frequent occlusions and high visual similarity between individuals. This work aimed to develop and evaluate a complete detection and tracking methodology for this specific scenario. The methodology involved creating annotated datasets, fine-tuning a YOLO11x object detector on custom data, and a comparative evaluation of four state-ofthe-art tracking algorithms (ByteTrack, BoT-SORT, HybridSORT, Deep OC-SORT). The finetuned YOLO11x detector achieved robust performance on the test set (mAP@50-95 = 86.9%). In the tracking evaluation, BoT-SORT yielded the highest overall accuracy (MOTA = 87.4%), while HybridSORT demonstrated the best identity consistency (IDF1 = 67.3%, lowest IDSW), revealing a trade-off between these metrics. It is concluded that combining the YOLO11x detector with either BoT-SORT or HybridSORT offers effective solutions for the problem, with the choice depending on the final application’s priority (overall tracking coverage vs. identity fidelity). This work contributes a detailed comparative analysis and a validated methodology for top-down MOT of dogs.