Source-linked AI summary
Privacy-Preserving Dataset Curation for Kuala Lumpur Urban Traffic: Grounded Vision-Language Detection with Spatial Vehicle-Context Filtering
Mohammed Abdul Al Arafat Tanzin, Rudzidatul Akmam Dziyauddin
TL;DR
Privacy-sensitive faces and vehicle plates make open urban-traffic imagery difficult to share, especially in challenging Kuala Lumpur scenes. The paper combines Grounding DINO with spatial vehicle-context filtering and anonymizes targeted regions in approximately 95% of 1,266 frames.
Problem
Open-sourcing urban traffic imagery requires anonymizing faces and vehicle registration plates to address legal and ethical concerns around PII exposure.
Method
The framework combines Grounding DINO open-set detection with a Spatial Vehicle ROI Containment Engine that validates license plates within vehicle boundaries.
Results
Approximately 95% of 1,266 examined frames successfully anonymized targeted privacy-sensitive regions.
Takeaways & Limitations
The framework supports auditable privacy-aware curation of complex Kuala Lumpur urban traffic imagery while preserving surrounding environmental context.
Takeaways & Limitations
The approach remains limited for distant or highly rotated motorcycle plates smaller than 16 × 16 pixels and for environmental false positives.
Abstract
from arXiv · showhide
The rapid advancement of intelligent transportation systems and autonomous driving relies heavily on multi-modal urban traffic datasets. However, curating high-fidelity video imagery in complex tropical urban environments---specifically Kuala Lumpur, Malaysia---presents severe challenges for Personally Identifiable Information (PII) anonymization due to high motorcycle density, dark acrylic license plates, dynamic camera tilt, and extreme tropical glare. We propose an automated anonymization framework tailored for the Kuala Lumpur Road Dataset, captured via a mobile cycling platform at 2 FPS. We document how legacy Haar cascades and YOLOv8 fail under these conditions---generating false positives on background elements while missing rotated or occluded targets. Our architecture resolves this by integrating Grounding DINO---a zero-shot open-set vision-language transformer---with a novel Spatial Vehicle Region of Interest (ROI) Containment Engine. By requiring license plate centroids to reside within validated vehicle boundaries, the pipeline suppresses environmental false positives while automatically obfuscating faces, heads, and license plates. An initial evaluation on 1,266 frames demonstrates a $\sim$95\% success rate, with remaining failures restricted to small, heavily occluded, oblique, or ambiguous targets. Coupled with temporal persistence mechanisms and an automated quality-control auditor, the framework minimizes privacy-related false negatives while preserving scene context for downstream vision tasks. While formal legal compliance depends on broader governance procedures, this publicly available pipeline and demonstration notebook provide an auditable preprocessing stage for privacy-aware dataset curation.
1. Introduction
The introduction frames privacy-preserving curation as necessary for reproducible urban-traffic research, while emphasizing that Kuala Lumpur imagery exposes failure modes for transferred anonymization systems. It presents a language-grounded detection and spatial vehicle-context filtering approach for anonymizing faces, heads, and registration plates.
- Motivation: Open-sourcing driving sequences supports reproducibility but exposes faces and vehicle registration plates as legally and ethically sensitive PII.The paper identifies PII exposure as a central challenge in developing ADAS and autonomous-navigation perception stacks.
- Dataset context: Kuala Lumpur traffic imagery challenges Western-benchmark anonymization architectures when captured from an iPhone 14 Pro mounted on a mobile bicycle platform.The dataset was collected in urban traffic in Kuala Lumpur, Malaysia, with video sampled at 2 frames per second.
- Approach: The study combines language-grounded object detection with spatial vehicle-context filtering to improve automated PII anonymization in challenging urban traffic imagery.The engineering evolution and mathematical formulation are accompanied by an interactive Jupyter Notebook showing execution, prompt configurations, and visual outputs.
- Contributions: The proposed curation pipeline targets faces, heads, and vehicle registration plates in urban traffic imagery acquired from a mobile bicycle-mounted camera in Kuala Lumpur.Its design is specifically oriented toward privacy-preserving image curation in this collection setting.
- Contributions: The work investigates a language-grounded vision transformer for open-set detection under small scale, occlusion, oblique views, helmets, and complex urban backgrounds.These conditions are identified as challenges for detecting privacy-sensitive objects.
2. Environmental Challenges in Kuala Lumpur Traffic Imagery
Kuala Lumpur traffic imagery poses privacy-sensitive detection challenges because motorcycles, helmets, and dynamic cycling-camera motion complicate facial visibility and object localization. Camera pitch and roll variations further degrade axis-aligned bounding-box representations used by closed-set CNNs.
- Motorcycle and Helmet Occlusions: Motorcycles are visually prominent, with compact dimensions, variable viewing angles, and helmets that partially or completely occlude facial regions.Headwear ranges from open-face to modular full-face helmets with dark visors, complicating conventional facial landmark detection.
- Dynamic Camera Motion: Dynamic cycling motion maps 3D world points to 2D pixels through camera intrinsics and rotation parameters for pitch, roll, and yaw.The projection is represented using intrinsic matrix K and dynamic rotation matrix R(θ, ϕ, ψ).
- Dynamic Camera Motion: Pitch and roll variations degrade the axis-aligned bounding-box representations used by closed-set CNN architectures.Camera tilt creates a mismatch between rotated image content and axis-aligned localization assumptions.
3. Evolution of Anonymization Architectures and Failure Modes
The anonymization architecture evolved from rigid Haar cascades through faster YOLOv8 models to Grounding DINO, addressing failures caused by dark plates, camera tilt, occlusion, small targets, and environmental false positives. Grounding DINO uses vision-language prompting to detect varied vehicles, plates, faces, and helmeted heads across orientations.
- Haar Cascade Failure Modes: Haar cascades failed on dark acrylic Malaysian plates, camera-induced perspective shearing, and helmeted riders that bypassed facial cascades.Their rigid intensity-gradient assumptions prevented reliable feature activation under these conditions.
- YOLOv8 Failure Modes: YOLOv8 improved feature-extraction speed but missed distant or highly rotated motorcycle plates smaller than 16 × 16 pixels.Its local receptive fields struggled with small pixel footprints.
- YOLOv8 Failure Modes: YOLOv8 also produced false plate detections on non-vehicle textures, road markings, speed bumps, and vehicle grilles.Deeper-stride feature maps blurred fine details and triggered environmental false positives.
- Grounding DINO Transition: Grounding DINO provided zero-shot open-set generalization by combining Swin Transformer visual features with BERT text-prompt features.The architecture integrates DINO with grounded visual-language pre-training.
- Grounding DINO Transition: Localized prompts enabled detection of custom acrylic plates and helmeted heads regardless of rotational orientation.Prompts covered vehicles, plates, faces, and heads, supporting multi-category anonymization targets.
4. Spatial Vehicle ROI Containment Engine
The Spatial Vehicle ROI Containment Engine suppresses environmental license-plate false positives by requiring candidate plate centroids to lie within validated vehicle boundaries. Validated detections are then expanded asymmetrically before obfuscation, while faces and heads receive uniform expansion.
- Spatial containment: The engine filters low-threshold transformer false positives by enforcing that valid license plates reside inside vehicle spatial boundaries.This spatial axiom targets detections on non-vehicle surfaces.
- Detection partitioning: Detections are partitioned into macro vehicles, candidate license plates, and face/head boxes before containment validation.Vehicle labels include vehicle, car, motorcycle, bus, and truck; plate labels include license plate and number plate.
- Containment validation: Candidate plate validity is determined from each plate centroid against dynamically expanded vehicle boundaries.Vehicle boxes expand by ϵ = 0.10 (10%) to accommodate tail-fenders and wide bumpers; plates without vehicles can pass when sp > 0.35.
- Obfuscation regions: Validated plates receive horizontal expansion αx = 0.25 (25%) and vertical expansion αy = 0.15 (15%), while faces receive uniform expansion αface = 0.20 (20%).The asymmetric padding addresses exposed plate margins or ear profiles left by tightly fitted predicted boxes.
5. Qualitative Analysis and Scenario Evaluation
The qualitative evaluation examines anonymization across challenging Kuala Lumpur traffic scenarios, including dense motorcycle traffic, background noise, and rotated plates under bicycle-camera motion. Grounding DINO and spatial vehicle containment address missed or false detections while preserving environmental context and covering angled plates.
- Scenario 1: Dense Motorcycle Traffic and Helmeted Riders: Dense motorcycle traffic defeats conventional face and plate detectors because helmets, rear-facing riders, and small oblique plates obscure targets.Grounding DINO detects prompted head, helmet, and plate targets, while the Spatial Vehicle ROI Containment Engine validates candidate plate centroids within expanded vehicle boundaries.
- Scenario 2: Background Noise Rejection: Candidate detections outside vehicle hulls are discarded, suppressing background false positives while preserving surrounding environmental context.This containment step evaluates each candidate plate centroid against an expanded vehicle boundary.
- Scenario 3: Dynamic Bicycle Camera Roll and Angled Plates: 25% horizontal mask expansion guarantees full character coverage for plates rotated by dynamic bicycle-camera pitch and roll.Grounding DINO cross-modal attention attends to global plate boundaries regardless of axis alignment.
- Qualitative Results: Figure 2 compares raw extracted frames with obfuscated outputs across helmeted motorcyclists, background-noise rejection, and rotated plates under sensor motion.The figure organizes these cases as rows (a), (b), and (c), respectively.
6. Temporal Persistence and Quality Control Auditing
The framework bridges single-frame detection dropouts through temporal interpolation and automates quality-control auditing with spatial vehicle-plate and person-face consistency rules. Passed frames are separated from flagged cases, with audit summaries preserving a defensible review trail.
- Temporal Persistence: Temporal linear interpolation restores missing target boxes when detections exist in the preceding and following frames of 2 FPS sequences.For It−1, It, It+1, the missing box at t is interpolated from bt−1 and bt+1.
- Quality Control Auditing: An automated context-aware auditor applies spatial consistency checks using YOLOv8l to eliminate manual review across thousands of frames.The auditor checks whether detected vehicles and pedestrians or riders contain corresponding validated privacy masks.
- Quality Control Auditing: 3000 pixels^2 is the vehicle-area threshold for flagging vehicles without validated plate masks.A vehicle box bv is flagged when its area is ≥3000 pixels^2 and M ∩ bv = ∅.
- Quality Control Auditing: 40 pixels is the pedestrian-or-rider height threshold for flagging boxes without face masks.A person box bperson is flagged when its height is ≥40 pixels and M ∩ bperson = ∅.
- Quality Control Auditing: Passed frames route to high_confidence/, while flagged frames move to review_required/ with qc_summary_report.json for an automated audit trail.The workflow reduces manual review requirements and supports defensible privacy-compliance auditing under frameworks including Malaysian PDPA 2010 and GDPR.
7. Conclusion and Code Availability
The conclusion emphasizes that Kuala Lumpur’s complex tropical traffic requires specialized vision pipelines beyond traditional rigid detectors and standard CNNs. It also provides the complete anonymization code, ROI engine, and demonstration notebook as open-source resources.
- Conclusion: Kuala Lumpur’s tropical traffic imagery requires specialized computer vision pipelines for vehicle modifications, dense motorcycle traffic, and dynamic sensor motion.These environmental and capture conditions motivate tailored dataset-curation methods.
- Conclusion: Traditional rigid object detectors and standard CNNs are insufficient for Kuala Lumpur traffic imagery’s extreme variability.The conclusion explicitly identifies these approaches as inadequate in this setting.
- Conclusion: The framework pairs Grounding DINO vision-language transformers with a Spatial Vehicle ROI Containment Engine.This pairing is presented as the framework’s central architectural approach.
- Code Availability: The complete source code, spatial ROI engine implementation, and interactive demonstration notebook are open-sourced for reproducible research and open science.The materials are available at the cited GitHub repository.