Skip to content

Computer Vision Counted Black-headed Gulls in Kunming with 91.6% Average Accuracy

Black-headed gull (Chroicocephalus ridibundus) in flight
Research · Other research

A computer-vision system tested in Kunming, China, counted black-headed gulls (Chroicocephalus ridibundus) in video with 91.6% average accuracy compared with manual counts by experienced observers. The peer-reviewed study, published on 5 August 2026, combined a modified YOLOv8 detector with the ByteTrack multi-object tracker.

The task is difficult because flying gulls can appear small in the image, overlap in dense groups and disappear briefly behind other birds. The detector identifies gulls in each frame, while ByteTrack links those detections through time and tries to preserve the same identity for the same bird. The counting step records each tracking identity once, reducing repeated counts of the same individual.

Three Kunming Sites Supplied the Test Material

Researchers recorded gulls at Cuihu Park, Daguan Park and the Haiceng Dam area between October 2024 and March 2026. They collected 894 still images and 38 videos lasting a total of three hours and 51 minutes. Still images and video frames produced a detection dataset of 2,170 images, while 20 videos were manually annotated with individual trajectories for the tracking tests.

Counts Stayed Close to Manual Review

With ByteTrack, the system achieved 89.7% multiple-object tracking accuracy, and ByteTrack gave the best overall tracking result among the five trackers compared with the same detector. For the intended monitoring task, the main result was the final count: automated counting reached 91.6% accuracy against manual counting across the 20 test sequences. This metric was calculated from the summed absolute differences between automated and manual counts, so the combined counting error corresponded to about 8.4% of the total manual count; it does not mean that 91.6% of individual gulls were independently identified correctly.

The result describes performance on the recorded Kunming material rather than a citywide estimate of gull abundance or a population trend. The authors also report lower robustness under extreme lighting and when image quality is poor. The measured processing speed of 21.4 frames per second came from a workstation with an RTX 3080 graphics processor, so further optimization and broader testing are needed before similar performance can be assumed on smaller field devices or in more varied environments.

The study therefore demonstrates a practical route from bird detection to individual tracking and automated counting in video. Its main contribution is not a new estimate of how many black-headed gulls live in Kunming, but a tested method for producing automated video counts under the conditions included in the study.

Sources and image credit

  1. An Automated Population Monitoring Framework for Larus ridibundus in Kunming City Based on Improved YOLOv8 and ByteTrackPeer-reviewed computer-vision study using 2,170 images and 38 videos from three Kunming sites; 20 videos were manually annotated for tracking and counting evaluation.

Image: © Giles Laurent, gileslaurent.com — wild black-headed gull (Chroicocephalus ridibundus) in flight at Lake Geneva, Switzerland, 30 December 2021; Wikimedia Commons, CC BY-SA 4.0. Representative species image; not from the Kunming study. · CC BY-SA 4.0 · Source

About this content: This story was produced with AI assistance within an editorial workflow developed by Wildlife Vagabond. Editorial responsibility remains with Wildlife Vagabond.How AI is used

Other research

SPECIES IN THIS STORY

Species in this story

Black-headed Gull Chroicocephalus ridibundus Explore species

Independent research and conservation news archive

Did you find this information useful?

Wildlife Vagabond is independently built and maintained. Voluntary support helps cover source verification, hosting and continued work on research and conservation news.

Support Wildlife Vagabond

The news archive will remain freely available.