A wildlife behaviour detection model: a case study on the red-legged seriema

Object detection models based on deep learning are increasingly used to reduce the manual workload associated with the processing of video recordings in behavioural research, but reproducible workflows describing how ecologists can build such models themselves remain scarce. Here, we present a step-by-step workflow for constructing a species-specific detection model and apply it as a case study to the red-legged seriema (Cariama cristata), a widely distributed but poorly studied Neotropical bird. Using 60 videos recorded with camera traps as well as non-stop recording cameras at the campus of the Universidade Federal de Vicosa (Florestal, Minas Gerais, Brazil) between 2023 and 2026, we extracted and annotated 480 images to train a YOLO26-nano detection model. On the validation set, the model achieved a precision of 0.934 and a recall of 0.786 (mAP@50 = 0.870), while performance on the independent test set was slightly lower, with a precision of 0.865 and a recall of 0.740 (mAP@50 = 0.740). When applied to unseen footage, the model processed video at roughly 4.6 times real-time speed on a standard CPU and correctly classified all four test videos for the presence or absence of seriemas. Most missed detections involved seriemas that were distant, partially out of frame, or recorded with the lower-quality non-stop cameras. These results showed that a usable detection model can be built from a comparatively small curated dataset, and the accompanying workflow and code are intended to lower the barrier for ecologists seeking to apply object detection to their own study species.