Problem
A safety incident such as crate toppling
occurs due to SOP violations. The
violation usually goes unnoticed until a
customer files a complaint days later.
To identify the root cause, supervisors
must watch hours of CCTV footage. This
is a tedious process that consumes up
to 8 hours
Under intense pressure to respond to
customers promptly, supervisors must
interview staff and draft formal reports.
The entire reactive resolution process
typically drags on for two full days
before the case can be resolved
"How do we automate SOP violation detection for DSV warehouse supervisors to improve violation visibility and response time as well as reduce manual effort in tracing back accidents?"
Our Solution
Hardware
Physical shell, embedded compute with Rasberry Pi and two Raspberry Pi cameras and HAILO accelerator, and mounting mechanism for warehouse deployment.
Software
Synthetic data generation, annotation, training a YOLO-based OBB model to detect crate stacking violations using rule-based algorithms. Detected violations trigger instant WhatsApp notifications to supervisors.
User Portal
Presents detected violations in a clear, user-friendly interface, allowing easy access and reviewing of incidents. An integrated Power BI dashboard provides analytics on warehouse performance based on recorded violations.
System Architecture
1
Camera
Capture
2
On-device
Inference
3
Ruleset
Trigger
4
Event
Recording
5
Notification
WhatsApp Alert
6
Dashboard
Review
7
Reporting
Closure
Software Architecture Diagram
Hardware Design Iterations
Iteration 1
The first hardware version was a bare Raspberry Pi with the camera module held in place by blue tack adhesive and a mini tripod stand that acted as a mount for the box-on-box stacking camera. Moreover, the whole mount was set up by a clamp which was functional for early testing but not suitable for deployment.
Drawing
Render
Actual
Iteration 2
Version 2 moved toward a formalized 3D-printed shell with integrated cable routing, a proper mounting bracket, and space for the HAILO accelerator. This made the system stable enough for real-world showcase and deployment.
Drawing
Render
Actual
Final Deployed Prototype
Closeup of Prototype
Set up at site
Software Design Iterations
SafeSight improved through multiple rounds of iteration across data, inference performance, and hardware deployment.
Data Iteration
The team initially faced a significant lack of warehouse CCTV footage capturing the target violations. Hence, we were unable to start working on our Computer Vision algorithms. To address this, synthetic data was generated early on using Google’s Nano Banana for images and Google Veo for videos.
1
Creating Camera Angles
As we lacked images from the required camera angles—top view for detecting back-to-back stacking and door view for box-on-box stacking—we used NanoBanana to generate synthetic data for these specific perspectives.
2
Generating Violation Images and Videos
We then used NanoBanana to generate synthetic violation data of back to back and box on box crate stackings, producing images for training our YOLO-based OBB models and videos for testing.
3
Refining Synthetic and Real Data
Synthetic data alone did not capture the full range of lighting conditions or the specific packaging used at DSV, resulting in poor bounding box performance across different crate types. To address this, we used our first prototype to collect real images and augment the training set. Combining live and synthetic data improved the model’s bounding box accuracy.
Data Annotation
To annotate the raw dataset, the platform Roboflow was used, leveraging its SAM 3 segmentation model to assist in labeling objects. This enabled efficient identification and outlining of key elements such as crates and forklifts through visual segmentation. The model helped automate much of the annotation process, reducing manual effort while maintaining consistency across the dataset, which ultimately improved the quality of training data for our YOLO-based OBB model.
Annotated Data
Annotated Data
Model Iteration
We then applied our YOLO-based OBB models using the refined dataset and implemented rule-based logic to detect violations. However, performance was limited on the Raspberry Pi, requiring further optimization due to reduced inference capability.
Box-on-Box Stacking
Back-to-Back Stacking
1
Version 1
Early edge deployment encountered latency constraints on the Raspberry Pi alone, running at only 3-5 FPS which made the system impractical for real-time monitoring.
2
HAILO Accelerator
Introducing the HAILO-8 AI accelerator module improved inference performance to 18-20 FPS with 10x lower latency, making the system responsive enough for live violation detection.
3
Improved OBB Model
The model still struggled to accurately predict OBB boxes, requiring further refinement of the training dataset. Performance was ultimately improved by adopting a more advanced YOLO OBB model from Data Iteration 3, leading to better inference results.
User Portal Iterations
The user portal acts as the operational hub for SafeSight after an event has been detected.
1
Initial Rough Mockup
An initial mockup was made based on user requirements with the base pages including a incident grid and an incident drawer to review specific incidents.
2
Figma Prototype
The mockup was then translated into a mid-fidelity Figma Prototype.
3
Final User Portal
The whole user experience was later revamped to make the whole incident reviewing experience conclusive and seamless with the addition of a settings page and an integrated PowerBI Dashboard.
Integrated PowerBI Dashboard Iterations
The Power BI dashboard provides warehouse supervisors with an analytics view of violations detected by SafeSight—categorized by time, location, and type—offering actionable insights into overall warehouse performance.
1
Iteration 1
The first iteration lacked visual polish and included several unnecessary fields that were not needed for our proof of concept.
2
Revised Dashboard
The revised version was refined to better meet the needs of DSV supervisors and redesigned with a cleaner, more modern interface.
3
Demo Dashboard
How it Works
Flow A
Real-Time Responses
1
Camera Monitors Activity
At night or during regular operations, the SafeSight device continuously captures video of the warehouse floor from its mounted position.
2
Cameras Detect Violation
On-device inference identifies SOP violations in real time — such as Back-to-Back Crate Stacking with bounding boxes, confidence scores, and tracking IDs.
3
System records relevant footage
Once a violation has been detected, the system automatically records a clip of the incident with timestamps for later review.
4
WhatsApp notification sent
The supervisor receives an immediate WhatsApp message with the violation type, location, time, and a thumbnail of the incident.
Flow B
Review and Reporting
1
Open User Portal
Warehouse supervisors access the SafeSight User Portal to see a full table of historical violation events, filtered by date, status, and type.
2
Review events and footage
Each event opens a detail drawer with video preview, incident summary, timestamps, and the ability to manually edit incident information.
3
Generate Reports and Close Cases
Completed reviews can be exported as formal incident reports with important information such as violation type, date and location etc.
Poster
System Impact
Operational Costs Reduced
Drives cost savings by avoiding lost productivity, product replacement and recrating costs with ~1.2 month ROI
*Assumes 20% accident reduction via WhatsApp, accounting for delays
Response Time Reduced
Reduces response time from hours to minutes via automated trace-back
*~5 minutes trace-back based on user testing
Carbon Emissions Reduced
Reduces emissions by minimising re-crating from fewer incidents
*Assumes average 1.89 metres cube crate size and 15 accidents/year
User Testimonials
Acknowledgements
This project would not have been possible without the guidance and support of several individuals. We are incredibly grateful to our industry mentor, Louise, for her time and expertise as well as Professor Thomas for his mentorship and feedback. A special thanks goes to Jia Ren, one of the warehouse supervisors, for his practical help on the ground.
