Enhance Drone Detection, Eliminate Bird False-Positives

Here’s a Freelancer project post focused on the bird-versus-drone problem: **Title: Improve Image-Only Drone Detection to Reduce Bird False Positives on Raspberry Pi + Hailo-10H** I have an existing Raspberry Pi camera pipeline running a single-class (“Drone”) YOLO model on a Hailo-10H. Birds, especially at a distance, may be detected as drones. I need a computer-vision engineer to develop and evaluate an image-only solution that reduces these false positives without losing too many real drone detections. The camera is fixed to the drone body. The current model is available as a compiled HEF; availability of its original training weights and dataset needs to be confirmed. Development must go in a new `DetectionV2` folder. Existing Detection, camera, telemetry, and interception code must remain unchanged. **Scope** - Establish a baseline using bird-only, drone-only, and bird-crossing-drone video. - Propose and test a practical approach, starting with a one-class drone model trained with bird footage as negative examples. - Evaluate whether information across video frames improves discrimination after accounting for camera movement. - Return “unknown” when distant images do not contain enough evidence to identify the object. - Deploy and benchmark the selected model or processing pipeline on Raspberry Pi + Hailo-10H. **Deliverables** - Reproducible training and evaluation code, dataset/labeling requirements, and model files, including a compatible HEF if a new model is trained. - DetectionV2 integration and tests, with no changes to the existing services. - Results on a held-out video set: bird false detections, confirmed false drone tracks, drone recall, decision delay, and device latency. - A short report explaining remaining failure cases, especially very small distant objects. This is a **detection and classification project only**. Flight control, collision avoidance, and interception guidance are outside scope. Please describe relevant drone-versus-bird video work, Hailo deployment experience, the data you would need, and how you would measure success. Do not promise perfect classification of objects that occupy only a few pixels.