Privacy-Preserving Dataset Curation for Kuala Lumpur Urban Traffic

An automated anonymization framework for the Kuala Lumpur Road Dataset is detailed in a new arXiv preprint (2608.14724). This addresses significant issues in managing high-quality video imagery essential for intelligent transportation systems and autonomous driving within complex tropical urban settings. The dataset, recorded at 2 FPS using a mobile cycling platform, struggles with anonymizing Personally Identifiable Information (PII) due to factors like high motorcycle traffic, dark acrylic license plates, variable camera angles, and intense tropical glare. The authors highlight that traditional Haar cascades and YOLOv8 are ineffective in these scenarios, resulting in false positives and overlooking rotated or occluded targets. Their solution combines Grounding DINO, a zero-shot open-set vision-language transformer, with a new Spatial Vehicle Region of Interest (ROI) Containment technique to enhance detection accuracy. This framework is specifically designed for Kuala Lumpur, Malaysia, promoting the secure sharing of urban traffic data while ensuring privacy. The paper is classified as a cross-type announcement and is accessible on arXiv.