AI-Powered Vehicle State Detection — Analytico Tech Case Study
Home/ Case Studies/ AI Vehicle State Detection
Computer Vision · AI Automotive Analytics

When cameras see a car park, most systems see nothing useful

Operators of parking facilities, automotive service yards, and logistics forecourts lose revenue and security oversight every day. Not because they lack cameras, but because those cameras cannot interpret what they see. We built an AI pipeline that changes that.

DomainAutomotive Intelligence
IndustrySmart Parking · Security · Fleet
DeliveryAI Model + Inference Pipeline
GeographyGlobal
StatusIn Development
The Business Problem

Cameras everywhere, actionable insight nowhere

Modern parking facilities, automotive service yards, and logistics compounds are covered in CCTV cameras. Yet almost all of those cameras serve a single purpose: passive recording for post-incident review. They cannot tell an operator that a vehicle has been left with its boot open in a secure bay. They cannot flag a car that is partially blocking an entry lane. They cannot help a parking management system understand which spaces are actually usable versus merely unoccupied on paper.

The gap is not cameras. It is intelligence at the point of capture. Operators are forced to rely on manual patrols, expensive sensor infrastructure embedded in tarmac, or simply accept the operational blind spots entirely.

🚗
Occluded vehicles go undetectedPartially hidden cars behind pillars or other vehicles cause miscounts and security gaps. Standard detection fails here.
🔓
Open boot: an unresolved security alertA vehicle left with its boot open in a secured forecourt is a security risk. No camera system was flagging it automatically.
🔩
Part-level visibility is missingDamage assessment, parts audit, and QA workflows in automotive service yards need granular part detection, not just "there is a car here."
💸
Existing solutions are costly to integrateEmbedded ground sensors and laser arrays require civil works. Camera-only AI that actually works at this granularity did not exist off the shelf.
How We Solved It

A purpose-built computer vision pipeline, trained on the real world

We designed and trained a multi-class object detection system from the ground up, tuned specifically for the conditions that break off-the-shelf models: partial occlusion, cluttered multi-vehicle scenes, wide variation in camera angles, and poor or mixed lighting.

1
Understanding the real operating environment
Before any training, we audited how cameras are actually deployed in parking and automotive service contexts: angles, distances, occlusion patterns, lighting conditions. This framed the dataset requirements rather than relying on clean benchmark imagery that would not reflect real deployments.
2
Curating and annotating a purpose-specific dataset
We built a multi-class annotated dataset covering four distinct detection categories: full vehicles, occluded vehicles, open car boots, and individual car parts such as wheels, doors and bumpers. Systematic augmentation including rotation, blur, brightness variation and mosaic tiling was applied to extend coverage of rare scenarios like open-boot edge cases.
RoboflowCustom annotationAugmentation pipeline
3
Training and iterating a YOLO-based detection model
We selected YOLOv8 for its single-stage speed-accuracy tradeoff and pre-trained COCO weights, giving the model strong initial vehicle feature representations to fine-tune from. Multiple training rounds with hyperparameter sweeps across learning rate, batch size and IoU thresholds optimised validation mAP across all four classes.
YOLOv8PythonGoogle Colab GPUHyperparameter sweeps
4
Building a lightweight, deployable inference pipeline
The trained model was wrapped in an OpenCV-based Python pipeline capable of running on both static images and live video frames. The pipeline outputs bounding boxes, class labels and confidence scores per detection, formatted to feed directly into parking management dashboards, security alert systems or automotive QA workflows.
OpenCVPython inferenceNMS per-class tuningBounding box rendering
5
Evaluating rigorously across realistic test conditions
Model performance was benchmarked across three distinct test sets: clean single-vehicle images, occluded multi-vehicle scenes, and dense cluttered parking lots. Per-class mAP, precision, recall and F1 were tracked across iterations. The architecture is dataset-scalable. Adding new vehicle types or states requires additional annotated data and a retraining run, with no changes to the pipeline itself.
mAP@50Precision / Recall / F1Confusion matrix3-dataset benchmark
System Architecture
Detection pipeline: input to output
📷
Camera / Image
Static · Video · Feed
→
⚙️
Preprocessing
OpenCV · Resize
→
🧠
YOLOv8
GPU Inference
→
📊
Detection Output
BBox · Class · Conf.
Output classes
🚗
Full Vehicle
Class 0
🚘
Occluded Vehicle
Class 1
🚪
Open Boot
Class 2
🔩
Car Parts
Class 3+
MLOps & training stack
1
Data collection
RoboflowAnnotationVersioningAugmentation
2
Model training
YOLOv8Google ColabPythonGPU compute
3
Inference pipeline
OpenCVPythonNMS tuningBBox rendering
4
Evaluation
mAP@50PrecisionRecallF1Confusion matrix
5
Deployment path
ONNX exportTensorRTEdge hardwareREST API
Results & Business Impact

From passive recording to real-time operational intelligence

4
Detection classes
30+
FPS on GPU
mAP 50+
Validation accuracy
3
Test environments
🔓
Security alerting becomes automatic
Open boot detection in secured areas can trigger immediate alerts without human monitoring. This was a gap no camera system previously addressed without manual patrol schedules.
🚘
Occluded vehicles no longer invisible
The model reliably detects partially hidden vehicles in cluttered multi-car scenes. This is the exact failure mode of off-the-shelf detection that operators previously had no answer to.
🔩
Parts-level insight for service operations
Automotive service yards gain the ability to assess vehicle state at component level, covering doors, bumpers and wheels, supporting damage documentation, QA workflows and parts audit without manual inspection.
⚡
Real-time, no infrastructure overhaul
Running at 30+ FPS on standard GPU hardware, the pipeline integrates with existing CCTV infrastructure. No ground sensors, no civil works and no hardware replacement required.
📈
Built to scale, not a one-off model
The pipeline architecture is class-agnostic. Adding new detection categories, whether vehicle types, new states or specific makes, requires only additional annotated data and a retraining run. The same pipeline has a clear path to edge deployment via ONNX export and TensorRT quantisation, making it viable for on-device inference at camera hardware level.
Technical Challenges

What makes this harder than standard object detection

⚡
Class imbalance: open boots are rareOpen-boot images are underrepresented in any natural dataset. Solved through targeted Roboflow augmentation and synthesis of edge-case samples to prevent the model from ignoring low-frequency classes.
⚡
High bounding box overlap in dense scenesCluttered parking lots produce heavily overlapping detections. Per-class NMS threshold tuning was critical to suppress false duplicates without losing genuine detections.
⚡
Small object detection at scaleIndividual car parts such as wheels, bumpers and door handles are small relative to full-frame context. Higher-resolution input tiling and mosaic augmentation addressed this without sacrificing inference speed.
⚡
Edge deployment constraintsCustomers cannot run cloud GPU inference continuously. ONNX export and TensorRT quantisation provide a path to on-device inference on camera hardware without accuracy sacrifice.
Key Technical Decisions

Why we built it this way

YOLOv8 over two-stage detectors
Faster R-CNN and similar two-stage architectures offer marginally higher accuracy on clean datasets but are far too slow for real-time video inference. YOLOv8's speed-mAP tradeoff was the right call for a deployment target of live CCTV feeds, with pre-trained COCO weights giving strong base vehicle representations to fine-tune from.
Roboflow for dataset management
Annotation, versioning, augmentation and export format management in a single platform is essential when iterating rapidly on dataset quality. Reproducible experiments, meaning the ability to roll back to a previous dataset version, proved critical during occlusion optimisation cycles.
OpenCV inference pipeline
A heavyweight ML framework dependency in the inference layer creates deployment friction. OpenCV is battle-tested, memory-efficient and runs on edge hardware. The detection pipeline can be deployed anywhere Python runs, including embedded systems attached to camera hardware.
ONNX + TensorRT export path
Training in PyTorch and exporting to ONNX decouples training infrastructure from deployment targets entirely. TensorRT quantisation then enables the same model to run at near-real-time speeds on modest edge GPU hardware, which is the target for on-premise automotive deployments.
Tech Stack
YOLOv8PythonOpenCV RoboflowGoogle ColabNumPy ONNXTensorRTPyTorch
Metrics
🎯
4 Classes
Vehicle · Occluded · Boot · Parts
⚡
30+ FPS
Real-time on GPU hardware
📊
mAP 50+
Validation dataset
🔄
Scalable
New classes via retraining only
🗂️
3 Envs
Clean · Occluded · Cluttered
Use Cases Enabled
Smart ParkingSecurity CCTV Automotive QADamage Assessment Fleet ManagementLogistics Yards

The problem is not a shortage of camera footage. It is a shortage of systems that can read it. We built the layer that turns video into decisions.

Analytico Tech — AI Engineering

Need AI that reads your environment?

We build custom computer vision systems, trained on your data and deployed on your infrastructure.

Discuss your project