Paper deep dive
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
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 8/18/2026, 4:24:12 AM
Summary
This paper presents an automated privacy-preserving anonymization framework for the Kuala Lumpur Road Dataset, addressing challenges in tropical urban traffic imagery such as high motorcycle density, non-standard license plates, and dynamic camera motion. The proposed architecture integrates Grounding DINO, a zero-shot vision-language transformer, with a novel Spatial Vehicle Region of Interest (ROI) Containment Engine to suppress false positives and accurately detect faces, heads, and license plates. The system achieved a 95% success rate on an initial evaluation of 1,266 frames, utilizing temporal persistence and automated quality control to ensure robust PII anonymization while preserving scene context for downstream tasks.
Entities (9)
Relation Signals (7)
Grounding DINO → usedin → Kuala Lumpur Road Dataset Anonymization
confidence 95% · Our architecture resolves this by integrating Grounding DINO... with a novel Spatial Vehicle Region of Interest (ROI) Containment Engine.
Spatial Vehicle ROI Containment Engine → supresses → Environmental False Positives
confidence 94% · By requiring license plate centroids to reside within validated vehicle boundaries, the pipeline suppresses environmental false positives
Kuala Lumpur Road Dataset → collectedby → Mobile Cycling Platform
confidence 92% · Kuala Lumpur Road Dataset, captured via a mobile cycling platform at 2 FPS.
YOLOv8 → failsunder → Kuala Lumpur Traffic Conditions
confidence 90% · We document how legacy Haar cascades and YOLOv8 fail under these conditions
GDPR → regulates → Personally Identifiable Information
confidence 88% · Legislative frameworks, such as the European General Data Protection Regulation (GDPR)... strictly regulate the processing... of unblurred public imagery
Grounding DINO → usesbackbone → Swin Transformer
confidence 85% · Grounding DINO processes multi-scale visual features F_v extracted via a Swin Transformer backbone
Grounding DINO → usesbackbone → BERT
confidence 85% · text prompt features F_t extracted via BERT
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2608.14724v1
- Canonical: https://arxiv.org/abs/2608.14724v1
Trouble viewing inline? Open PDF directly →
Full Text
25,340 characters extracted from source content.
Expand or collapse full text
Privacy-Preserving Dataset Curation for Kuala Lumpur Urban Traffic: Grounded Vision-Language Detection with Spatial Vehicle-Context FilteringJournal: Computer Vision and Image Understanding Mohammed Abdul Al Arafat Tanzin, Email: mohammedabdul@graduate.utm.my Address: Faculty of Artificial Intelligence, Universiti Teknologi Malaysia, Kuala Lumpur, Malaysia Corresponding author: Corresponding author. Rudzidatul Akmam Dziyauddin Address: Faculty of Artificial Intelligence, Universiti Teknologi Malaysia, Kuala Lumpur, Malaysia Abstract 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 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. Keywords: Dataset Anonymization , Privacy Preservation , Vision-Language Models , Grounding DINO , Kuala Lumpur Road Dataset , Intelligent Transportation Systems 1 Introduction Developing robust perception stacks for Advanced Driver Assistance Systems (ADAS) and autonomous navigation requires diverse training footage captured across varied geographical and structural driving environments. Open-sourcing raw driving sequences is essential for scientific reproducibility; however, exposing Personally Identifiable Information (PII)—specifically human faces and vehicle registration plates—creates significant legal liabilities and ethical concerns. Legislative frameworks, such as the European General Data Protection Regulation (GDPR) and Malaysia’s Personal Data Protection Act 2010 (PDPA), strictly regulate the processing and public dissemination of unblurred public imagery [2]. While automated anonymization pipelines have been deployed effectively on Western driving benchmarks (such as nuScenes [1] and Waymo [6]), directly transferring these architectures to Southeast Asian traffic streams reveals severe failure modes. The Kuala Lumpur Road Dataset considered in this study was collected using an iPhone 14 Pro optical sensor mounted on a mobile bicycle platform while navigating through urban traffic in Kuala Lumpur, Malaysia. Video sequences were sampled at 2 frames per second (FPS) to generate image frames for dataset curation. The initial anonymization experiment reported in this paper comprises 1,266 extracted frames representing diverse traffic and environmental conditions. The mobile acquisition platform introduces time-varying pitch (θ), roll (ϕφ), vibration, and perspective changes, resulting in substantial variation in the apparent scale, orientation, and visibility of privacy-sensitive objects such as faces and vehicle registration plates. The dynamic motion profile of a cycling platform introduces time-varying pitch (θ), roll (ϕφ), and vibration, generating severe rotational variation and perspective shearing across image frames. In this paper, we detail the engineering evolution and mathematical formulation of the Kuala Lumpur Road Dataset Anonymizer11 1 Project Repository and Interactive Jupyter Notebook: https://github.com/TanzinAbdul/Kuala-Lumpur-Road-Dataset-Anonymizer. Readers are directed to the project repository’s interactive Jupyter Notebook (.ipynb) for step-by-step pipeline execution, cross-modal prompt configurations, and visual output demonstrations. We investigate how combining language-grounded object detection with spatial vehicle-context filtering can improve automated PII anonymization in challenging urban traffic imagery [4]. In an initial qualitative assessment of 1,266 extracted frames, the proposed pipeline successfully anonymized the targeted privacy-sensitive regions in approximately 95% of the examined frames, while the remaining cases revealed characteristic failure modes involving small objects, severe occlusion, extreme viewing angles, and ambiguous visual structures. The contributions of this work are summarized as follows: 1. We present a privacy-preserving image curation pipeline designed specifically for urban traffic imagery acquired from a mobile bicycle-mounted camera in Kuala Lumpur, with particular emphasis on faces, heads, and vehicle registration plates. 2. We investigate the use of a language-grounded vision transformer for open-set detection of privacy-sensitive objects under challenging conditions including small object scale, occlusion, oblique viewing angles, helmets, and complex urban backgrounds. 3. We introduce a Spatial Vehicle ROI Containment Engine that uses the spatial relationship between candidate license plates and detected vehicle regions to suppress detections that are inconsistent with the expected topology of a vehicle. 4. We incorporate temporal persistence and automated quality-control auditing to identify potential detection failures and reduce the probability of undetected privacy-sensitive objects remaining in the curated dataset. 5. We provide an initial qualitative evaluation on 1,266 extracted frames and document the observed success cases and failure modes to establish a reproducible basis for subsequent quantitative evaluation. 2 Environmental Challenges in Kuala Lumpur Traffic Imagery Southeast Asian traffic landscapes differ fundamentally from Western driving environments. The primary operational domain complexities encountered in Kuala Lumpur include: 1. High Motorcycle Presence and Helmet Occlusions: Motorcycles constitute a visually prominent component of urban traffic in Kuala Lumpur. Their compact dimensions, variable viewing angles, and frequent use of helmets create additional challenges for privacy-sensitive detection, particularly when facial regions are partially or completely occluded. Motorcyclists wear diverse headwear ranging from open-face to modular full-face helmets with dark visors, rendering conventional facial landmark detectors ineffective. 2. Custom Non-Standard Acrylic Plates: Unlike standardized metal reflective plates, Malaysian registration plates predominantly feature custom black acrylic backings with non-standard white typography, variable spacing, and non-conforming mounting profiles on modified vehicles. 3. Dynamic Sensor Tilt and Perspective Shearing: The mobile bicycle mount induces transient camera pitch and roll, causing extreme rotational angles (θ>30∘θ>30 ) for observed registration plates, breaking axis-aligned bounding box assumptions. 4. High-Contrast Background Visual Noise: Equatorial sunlight creates intense specular reflections, while complex street environments, speed bump patterns, and vehicle grilles frequently mimic alphanumeric character arrangements. To quantify these domain differences, Table 1 compares standard Western benchmark characteristics against the Kuala Lumpur road environment. Table 1: Domain comparison between Western benchmark datasets and Kuala Lumpur urban traffic imagery. Parameter Western Datasets Kuala Lumpur Traffic Impact on Detection Plate Material Standard Retro-Reflective Metal Custom Black Acrylic Low contrast; edge detectors fail Motorcycle Ratio <5%<5\% of traffic stream 30%−50%30\%-50\% of traffic stream Oblique rear mounting profiles Rider Headwear Infrequent / Unobstructed Helmets, Visors, Balaclavas Facial landmarks unavailable Camera Rig Rigid Vehicle Roof Mount Mobile Bicycle Mount High pitch, roll, and vibration Background Noise Uniform pavement textures High-contrast grilles & patterns False positives on background noise Mathematically, the image projection process under dynamic cycling motion is defined by mapping a 3D world point w=[Xw,Yw,Zw,1]TP_w=[X_w,Y_w,Z_w,1]^T to a 2D pixel coordinate =[u,v,1]Tp=[u,v,1]^T via camera intrinsic matrix K and dynamic rotation matrix (θ,ϕ,ψ)R(θ,φ,ψ): ∼[z(ψ)y(θ)x(ϕ);]wp [R_z(ψ)R_y(θ)R_x(φ)\;;\;t ]P_w (1) where pitch (θ) and roll (ϕφ) variations degrade axis-aligned bounding box representations used in closed-set CNN architectures. 3 Evolution of Anonymization Architectures and Failure Modes 3.1 Legacy Approaches: Haar Cascades Initial iterations utilized OpenCV Haar feature-based cascade classifiers [7]. Haar cascades compute feature values fH(x)f_H(x) across adjacent rectangular sub-windows using integral images: fH(x)=∑r∈Rwhitewr⋅I(r)−∑r∈Rblackwr⋅I(r)f_H(x)= _r∈ R_whitew_r· I(r)- _r∈ R_blackw_r· I(r) (2) Haar cascades completely failed on dark acrylic Malaysian plates due to rigid intensity gradient assumptions. Furthermore, perspective shearing caused by camera tilt prevented feature activation, while helmeted riders bypassed facial cascades entirely. 3.2 Closed-Set Deep CNNs: YOLOv8 Transitioning to deep single-stage convolutional networks improved feature extraction speeds. We employed YOLOv8 [3], which optimizes bounding box coordinates =[xc,yc,w,h]Tb=[x_c,y_c,w,h]^T and class distributions using a composite loss: ℒYOLO=λboxℒCIoU+λclsℒBCE+λdflℒDFLL_YOLO= _boxL_CIoU+ _clsL_BCE+ _dflL_DFL (3) Despite impressive inference throughput, fine-tuned YOLOv8 models exhibited critical domain limitations: 1. Local Grid Constraints: Local receptive fields failed to detect distant or highly rotated motorcycle plates occupying small pixel footprints (<16×16<16× 16 pixels). 2. Environmental False Positives: Feature maps at deeper strides blurred fine details, causing non-vehicle textures, road markings, speed bumps, and vehicle grilles to trigger false plate detections. 3.3 Proposed Architecture: Grounding DINO Vision-Language Transformer To achieve zero-shot open-set generalization, we integrated Grounding DINO [4], a transformer architecture marrying DINO with grounded visual-language pre-training. Grounding DINO processes multi-scale visual features vF_v extracted via a Swin Transformer backbone [5] and text prompt features tF_t extracted via BERT [2]. Cross-modal interaction is governed by bi-directional feature enhancement: v(l)=CrossAttention(v(l−1),t(l−1),t(l−1))F_v^(l)=CrossAttention(F_v^(l-1),F_t^(l-1),F_t^(l-1)) (4) t(l)=CrossAttention(t(l−1),v(l−1),v(l−1))F_t^(l)=CrossAttention(F_t^(l-1),F_v^(l-1),F_v^(l-1)) (5) Object queries are dynamically initialized by computing the vision-language alignment score matrix align=vtTS_align=F_vF_t^T. Prompting the transformer with localized natural language terms—“vehicle, car, motorcycle, bus, truck, license plate, number plate, human face, head”—enables multi-head self-attention across the frame, detecting custom acrylic plates and helmeted heads regardless of rotational orientation. 4 Spatial Vehicle ROI Containment Engine 4.1 Pipeline System Architecture Figure 1 illustrates the complete end-to-end framework, detailing text encoding, cross-modal vision-language fusion, macro/micro box partitioning, spatial containment filtering, dynamic obfuscation, and automated quality control auditing. The implementation details are fully executable via the project repository’s Jupyter Notebook (.ipynb). Figure 1: End-to-end architecture of the proposed Spatial-Semantic Anonymization Framework for the Kuala Lumpur Road Dataset. License plate candidates (P) undergo spatial containment validation against macro vehicle ROIs (V), while faces and helmeted heads (ℱF) bypass spatial filtering and route directly to the obfuscation module. 4.2 Containment Logic and Topology Formulation To eliminate environmental background false positives generated by low-threshold transformer detection on non-vehicle surfaces, we engineered the Spatial Vehicle ROI Containment Engine. The foundational spatial axiom dictates that a valid license plate must reside inside the spatial boundary of a vehicle. Given output detection set =(k,sk,ℓk)k=1KD=\(b_k,s_k, _k)\_k=1^K, detections are partitioned into three disjoint sets: =v∣(v,sv,ℓv)∈,ℓv∈vehicle, car, motorcycle, bus, truck =\b_v (b_v,s_v, _v) , _v∈\vehicle, car, motorcycle, bus, truck\\ (6) =(p,sp)∣(p,sp,ℓp)∈,ℓp∈license plate, number plate =\(b_p,s_p) (b_p,s_p, _p) , _p∈\license plate, number plate\\ (7) ℱ =f∣(f,sf,ℓf)∈,ℓf∈human face, head =\b_f (b_f,s_f, _f) , _f∈\human face, head\\ (8) For each candidate plate box p=[x1,p,y1,p,x2,p,y2,p]∈b_p=[x_1,p,y_1,p,x_2,p,y_2,p] , its geometric centroid p=(xc,p,yc,p)c_p=(x_c,p,y_c,p) is calculated as: xc,p=x1,p+x2,p2,yc,p=y1,p+y2,p2x_c,p= x_1,p+x_2,p2, y_c,p= y_1,p+y_2,p2 (9) To accommodate tail-fenders and wide vehicle bumpers, macro vehicle boxes v=[x1,v,y1,v,x2,v,y2,v]b_v=[x_1,v,y_1,v,x_2,v,y_2,v] are dynamically expanded by expansion factor ϵ=0.10ε=0.10 (10%): x~1,v x_1,v =max(0,x1,v−ϵwv),y~1,v=max(0,y1,v−ϵhv) = (0,x_1,v-ε w_v), y_1,v= (0,y_1,v-ε h_v) (10) x~2,v x_2,v =min(W,x2,v+ϵwv),y~2,v=min(H,y2,v+ϵhv) = (W,x_2,v+ε w_v), y_2,v= (H,y_2,v+ε h_v) (11) The containment indicator function valid(p)I_valid(b_p) evaluates candidate validity: valid(p)=1if ∃vϵ s.t. (x~1,v≤xc,p≤x~2,v∧y~1,v≤yc,p≤y~2,v)1else if (=∅∧sp>0.35)0otherwise (Discard as Non-Vehicle Background Noise)I_valid(b_p)= cases1&if _v^ε s.t. ( x_1,v≤ x_c,p≤ x_2,v y_1,v≤ y_c,p≤ y_2,v)\\ 1&else if (V= s_p>0.35)\\ 0&otherwise (Discard as Non-Vehicle Background Noise) cases (12) Algorithm 1 details the complete spatial containment filtering process. Algorithm 1 Spatial Vehicle ROI Containment Filtering Algorithm 1: Frame ∈ℝH×W×3I ^H× W× 3, Detection Set D, Expansion Tolerance ϵ=0.10ε=0.10 2: Validated Anonymization Mask Map ∈0,1H×WM∈\0,1\^H× W 3: Initialize ←H×WM 0^H× W 4: Partition →D→ Macro Vehicles V, Candidate Plates P, Faces/Heads ℱF 5: for each candidate plate (p,sp)∈(b_p,s_p) do 6: Compute centroid p←(x1,p+x2,p2,y1,p+y2,p2)c_p← ( x_1,p+x_2,p2, y_1,p+y_2,p2 ) 7: is_valid←Falseis\_valid 8: for each vehicle box v∈b_v do 9: Compute expanded boundary vϵ=[x~1,v,y~1,v,x~2,v,y~2,v]b_v^ε=[ x_1,v, y_1,v, x_2,v, y_2,v] 10: if x~1,v≤xc,p≤x~2,vandy~1,v≤yc,p≤y~2,v x_1,v≤ x_c,p≤ x_2,v\;and\; y_1,v≤ y_c,p≤ y_2,v then 11: is_valid←Trueis\_valid ; break 12: end if 13: end for 14: if is_validor(=∅andsp>0.35)is\_valid\;or\;(V= \;and\;s_p>0.35) then 15: Apply Asymmetric Padding Expansion to p→p∗b_p _p^* 16: Update Mask Map [p∗]←1M[b_p^*]← 1 17: end if 18: end for 19: for each face/head box f∈ℱb_f do 20: Expand f→f∗b_f _f^*; Update Mask Map [f∗]←1M[b_f^*]← 1 21: end for 22: return M 4.3 Asymmetric Mask Expansion and Dynamic Gaussian Blurring Raw predicted bounding boxes often fit tightly around central characters, leaving plate margins or ear profiles exposed. Validated boxes undergo asymmetric expansion: horizontal bounds expand by αx=0.25 _x=0.25 (25%) and vertical bounds by αy=0.15 _y=0.15 (15%) for plates, while faces undergo uniform expansion αface=0.20 _face=0.20 (20%). Dynamic Gaussian blurring is executed across expanded regions ROII_ROI using a 2D kernel G(x,y,σx,σy)G(x,y; _x, _y): G(x,y,σx,σy)=12πσxσyexp(−(x22σx2+y22σy2))G(x,y; _x, _y)= 12π _x _y (- ( x^22 _x^2+ y^22 _y^2 ) ) (13) Kernel dimensions (kw,kh)(k_w,k_h) scale dynamically with ROI resolution to prevent boundary artifacts: kw=max(3,2⌊wROI2⌋+1),kh=max(3,2⌊hROI2⌋+1)k_w= (3,2 w_ROI2 +1 ), k_h= (3,2 h_ROI2 +1 ) (14) 5 Qualitative Analysis and Scenario Evaluation Evaluating anonymization performance across complex Kuala Lumpur traffic imagery requires analyzing qualitative behavior across challenging operational scenarios encountered in the dataset. 5.1 Qualitative Analysis across Traffic Scenarios Figure 2 illustrates qualitative outputs generated by the proposed framework across key challenge scenarios in Kuala Lumpur traffic. (a) Scenario 1: High-density motorcycle stream with helmeted riders. (b) Scenario 2: Background false positive suppression via vehicle ROI containment. (c) Scenario 3: Angled vehicle plates under dynamic bicycle camera roll. Figure 2: Qualitative anonymization results on raw extracted frames (left) versus obfuscated outputs (right) from the Kuala Lumpur Road Dataset. Row (a): Helmeted motorcyclists and oblique tail plates. Row (b): Background noise rejection. Row (c): Rotated plates under sensor motion. 1. Scenario 1: Dense Motorcycle Traffic and Helmeted Riders: In raw imagery capturing dense lane-filtering motorcycles, conventional face detectors fail because riders wear full-face helmets or face away from the camera. Legacy plate detectors miss small, oblique tail-mounted plates. By prompting Grounding DINO [4] with “head, helmet, motorcycle license plate”, the pipeline captures helmeted rider profiles and angled rear plates simultaneously. 2. Scenario 2: High-Contrast Background Noise Suppression: Specular reflections and high-contrast street elements often trigger false positive activations in unconstrained object detectors. The Spatial Vehicle ROI Containment Engine evaluates the centroid pc_p of each candidate plate box and verifies whether it lies within an expanded vehicle boundary vϵb_v^ε. Detections outside vehicle hulls are discarded, preserving surrounding environmental context. 3. Scenario 3: Dynamic Bicycle Camera Roll and Angled Plates: Cycling motion generates variable pitch and roll (ϕ,θφ,θ), causing observed license plates to appear at acute rotational angles. Cross-modal attention maps in Grounding DINO attend to global plate boundaries regardless of axis alignment, while 25% horizontal mask expansion guarantees full character coverage. 6 Temporal Persistence and Quality Control Auditing 6.1 Temporal Persistence Tracking To bridge single-frame detection dropouts in extracted 2 FPS video sequences, the framework executes a temporal linear interpolation pass. For frame sequence t−1,t,t+1I_t-1,I_t,I_t+1, if a target box is detected at t−1t-1 (t−1b_t-1) and t+1t+1 (t+1b_t+1) but missing at frame t, the missing coordinate is interpolated: t=t−1+t+12b_t= b_t-1+b_t+12 (15) 6.2 Automated Quality Control (QC) Audit Architecture To eliminate manual review across thousands of frames, an automated context-aware auditor evaluates two spatial consistency rules using a large secondary context model (YOLOv8l [3]): 1. Vehicle-Plate Audit Rule: If a vehicle box vb_v occupies an area ≥3000 pixels2≥ 3000 pixels^2 but contains no validated plate mask (∩vϵ=∅M _v^ε= ), the frame is flagged. 2. Person-Face Audit Rule: If a pedestrian/rider box personb_person exceeds a height threshold ≥40 pixels≥ 40 pixels but contains no face mask (∩person=∅M _person= ), the frame is flagged. Passed frames route directly to high_confidence/, while flagged frames isolate into review_required/ alongside an automated audit summary (saved as qc_summary_report.json). This context audit reduces manual review requirements significantly, providing a defensible audit trail for privacy compliance under legal frameworks like the Malaysian PDPA 2010 and GDPR. 7 Conclusion and Code Availability Curating open-source urban traffic datasets in complex tropical environments requires specialized computer vision pipelines capable of handling non-standard vehicle modifications, high motorcycle density, and dynamic sensor motion. We demonstrated that traditional rigid object detectors and standard CNNs are insufficient for the extreme variability of Kuala Lumpur traffic imagery. By pairing Grounding DINO vision-language transformers with a Spatial Vehicle ROI Containment Engine, our framework achieves accurate anonymization of faces, helmeted heads, and custom license plates while suppressing environmental false positives. To support reproducible research and open science, the complete source code, spatial ROI engine implementation, and interactive demonstration notebook (.ipynb) are open-sourced at https://github.com/TanzinAbdul/Kuala-Lumpur-Road-Dataset-Anonymizer. References [1] H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q. Xu, A. Krishnan, Y. Pan, G. Baldan, and O. Beijbom (2020) NuScenes: a multimodal dataset for autonomous driving. External Links: 1903.11027, Link Cited by: §1. [2] J. Devlin, M. Chang, K. Lee, and K. Toutanova (2019) BERT: pre-training of deep bidirectional transformers for language understanding. External Links: 1810.04805, Link Cited by: §1, §3.3. [3] Ultralytics yolov8 External Links: Link Cited by: §3.2, §6.2. [4] S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Su, J. Zhu, and L. Zhang (2024) Grounding dino: marrying dino with grounded pre-training for open-set object detection. External Links: 2303.05499, Link Cited by: §1, §3.3, item 1. [5] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo (2021) Swin transformer: hierarchical vision transformer using shifted windows. External Links: 2103.14030, Link Cited by: §3.3. [6] P. Sun, H. Kretzschmar, X. Dotiwalla, A. Chouard, V. Patnaik, P. Tsui, J. Guo, Y. Zhou, Y. Chai, B. Caine, V. Vasudevan, W. Han, J. Ngiam, H. Zhao, A. Timofeev, S. Ettinger, M. Krivokon, A. Gao, A. Joshi, S. Zhao, S. Cheng, Y. Zhang, J. Shlens, Z. Chen, and D. Anguelov (2020) Scalability in perception for autonomous driving: waymo open dataset. External Links: 1912.04838, Link Cited by: §1. [7] P. Viola and M. Jones (2001) Rapid object detection using a boosted cascade of simple features. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, Vol. 1, p. I–I. External Links: Document Cited by: §3.1.