Cross-Domain Validation Framework for Thai Roadway Accident Recognition: Data, Trained Weights, and Code

Supporting materials for the paper "Frame-Level Accident Recognition via Detection Confidence Aggregation: A Cross-Domain Validation Framework for Thai Roadway Surveillance" submitted to Technologies (MDPI). Contents:1. Thai inference-pool metadata for 1,245 frames (23 positive accident events).2. Trained detector weights: YOLOv5su, YOLOv8n, YOLOv11n v6 across three random seeds (42, 7, 123) with focal loss, and YOLOv11n v7 fine-tuned on Chiang Mai footage.3. Python scripts: confidence-aggregation operators (max, mean, top-K, class-weighted), DeLong variance and bootstrap CI computation, Kullback-Leibler divergence on confidence histograms, per-platform inference benchmarks.4. Per-seed AUROC, DeLong 95% CI, and bootstrap 95% CI tables for all 24 metric configurations.5. Source-validation inference output for v6 seed-7 (412 frames, 162 positives) used in the cross-domain retention ratio.6. Inference timing on three platforms: NVIDIA RTX 5070 Ti (CUDA), Apple Silicon (MPS), and Intel i7-class CPU. Raw video footage is not redistributable under the source-feed licence. Researchers may request access through the corresponding author.