WildVision AI

Monitoring wildlife across hours of safari footage is slow and error-prone, and single-frame detections often misidentify fast-moving or partially visible animals.

• Manually reviewing long video feeds is time-consuming for park rangers.
• Single-frame detections produce false positives on occluded or blurred animals.
• Distinguishing visually similar species (e.g. antelopes) requires high accuracy.
• Processing full video at every frame is computationally expensive.

We trained a custom YOLOv11 model and wrapped it in a frame-sampling, voting-based pipeline exposed through both a web dashboard and a CLI.

• Custom YOLOv11 model trained on a Roboflow-annotated dataset of 12 species.
• Key-frame extraction to speed up processing without sacrificing accuracy.
• Voting mechanism aggregating detections across frames to cut false positives.
• FastAPI web dashboard, CLI tool, and RESTful API for flexible access.