Event-based EarlyDetection for YOLO Models : Comparison of Event-Handling Methodsand Real-Time Use in UAV’s

Sammanfattning

This thesis investigates whether the low-latency properties of event cameras can enable fast, real-time object detection. The study focuses on YOLO models, known for their rapid inference, and includes an examination of a recurrent YOLO vari- ant. Since YOLO is image-based, various event-to-image methods are explored. Findings indicate that non-recurrent models are faster, making them more suit- able for early detection. Effective accumulation methods achieve a mAP@50 of 0.55. Potential applications include UAVs, so the models are also tested on aerial event data. The findings from these tests reveals the need for UAV-specific train- ing to improve performance. Future work could explore shorter event accumu- lation intervals to increase fps, analyze trade-offs between model size, version, mAP, and FPS, and optimize the ReYOLOv8n training setup to enhance mAP.