DRGD: A Deterministic Color–Geometry–PnP Gate Detector for Reproducible CPU-Based Evaluation on Indoor and Outdoor Sequences

Abstract

This paper presents DRGD (Drone Racing Gate Detector), a modular classical computer-vision detector for single-gate localization in indoor and outdoor drone-gate sequences. Its contribution is system-level rather than operator-level: standard color, morphology, geometry, and PnP components are inte grated into a deterministic color–geometry–PnP pipeline with frozen domain configurations, split-disciplined evaluation, staged ablation, and explicit reporting of negative evidence. DRGD follows a two-stage classical detection structure: Stage 1 converts full-frame color evidence into mask-based color–geometry proposals, while Stage 2 applies geometric constraints, PnP plausibility validation, and top-K candidate selection before emitting the final bounding box. The detector is evaluated on a custom four-sequence 640×480 dataset comprising SQ3, SQ4, SQ5, and SQ7, using a split-disciplined CPU protocol with calibration, validation, configuration freezing, A0–A3 ablation, and a single-run final test. After indoor calibration, the frozen 1d9 indoor configuration achieved 0.953 F1 and 45.91 FPS on validation, and 0.905 F1 and 48.46 FPS on the indoor final test. The frozen 1d11 outdoor configuration achieved 0.891 F1 and 42.08 FPS on validation, but reached 0.474 F1 globally on the outdoor final test, mainly due to the lower SQ7 result. To provide an external quantitative reference, a YOLO11n baseline was fine-tuned on the same splits and evaluated with the same frame-level test protocol; it reached 0.998 indoor F1 and 0.840 outdoor F1, improving SQ7 recall while still producing false positives on gate-absent SQ7 frames. The results show that DRGD is accurate and real-time under controlled indoor conditions, remains usable on SQ5 outdoors, and identifies sequence-dependent outdoor robustness as the main direction for future embedded deployment.