Add CynapFleet zero-shot submission

[smammahdi]

Submission Metadata

Please fill out every field below. These values are used for benchmark review and leaderboard display.

method_name: Two-Pass Hybrid YOLO11x + Qwen3-VL Coarse-to-Fine CCTV Accident Detector
authors: Dipit Saha, Shah Mohammad Abdul Mannan, Mohammad Raihan Rashid, Ruwad Naswan, Ahnaf Tahmid
affiliation: CSE, Bangladesh University of Engineering and Technology (BUET)
paper_url: N/A
code_url: N/A
contact_email: mahdithe8@gmail.com
benchmark: zero-shot
short_description: Zero-shot CCTV accident detection pipeline using Qwen3-VL with a hierarchical two-pass coarse-to-fine design. Coarse pass uniformly samples 60 frames at 4 fps with embedded second-level timestamps, prompting the VLM to produce an initial accident time and collision-type hint. Fine pass deploys YOLO11x + BoT-SORT tracking at native FPS within a ±2 s refinement window, subsampling 24 key frames; each frame is annotated with per-vehicle bounding boxes overlaid visually and injected as normalized coordinate text, while the coarse type serves as a soft prior. Final accident time, spatial center, and collision type are extracted from the fine-pass VLM response.

Checklist

• I added exactly one submission CSV under submissions/zero-shot/
• The CSV follows the required ACCIDENT submission format
• The metadata block above is complete and accurate

[github-actions]

Scoring preview for zero-shot benchmark:

• Unified score: 0.5042
• Type accuracy: 0.5027
• Temporal @ sigma=1.0: 0.5486
• Spatial @ sigma=1.0: 0.4677

If this PR is merged, the leaderboard update workflow will open a follow-up PR with the public leaderboard change.