RIOS

Traditional POS systems show completed transactions, but they miss lost intent before checkout. RIOS was designed to make pre-purchase shelf behavior measurable and useful for decisions around placement, replenishment, pricing, and conversion analysis.

• Attention without purchase can reveal hidden conversion gaps.
• Strong shelf interaction may signal stock, price, or placement issues.
• Behavior signals give inventory and sales decisions more context.

RIOS was developed during the Applied Artificial Intelligence Innovation Challenge 2026 as a team project. The goal was to turn a broad retail analytics idea into a working full-stack AI system connecting computer vision, backend services, dashboards, and decision support.

• Built by Team Decipher during AIIC 2026.
• Worked as a 4-member team combining 3 ICT and 1 Business Students.
• Focused on delivering a real system flow rather than only a model demo.
• Delivered a working workflow across computer vision, IoT exploration, analytics dashboards, and AI-assisted reporting.

RIOS connects physical shelf activity with analytics and decision support through a full behavior-to-insight pipeline.

• Camera or IoT feed produces shelf-level visual input.
• YOLO detection converts video frames into behavior events.
• Behavior events are combined with sales and inventory data.
• Processed signals feed dashboards, scoring logic, and AI-assisted reports.

The vision dataset was collected manually in real retail environments to support behavior-level detection rather than generic object detection.

• Collected 11 source videos from shelf-facing retail scenarios.
• Annotated 10 videos using CVAT, with 1 video reserved for testing.
• Used video-level splitting to reduce leakage across similar scenes.
• Focused on behavior classes such as no interaction, touch, hold, and item removal.

The computer vision layer detects shelf-level interaction behaviors and converts them into usable behavior signals for the wider retail system.

• Uses an Ultralytics YOLO pipeline for custom retail behavior detection.
• Turns frame-level detections into behavior events such as touch, hold, and item removed.
• The model acts as the signal generator for the retail intelligence pipeline.

These test samples show the detection model running on held-out retail shelf footage, making the behavior detection output easier to understand visually.

The custom YOLO model was evaluated using standard detection metrics and visual diagnostics to understand class-level behavior performance.

• Trained for 50 epochs, with best observed performance around epoch 25.
• Visually distinct actions such as item removal were easier than subtle actions such as touch versus hold.
• Evaluation visuals were used to inspect class confusion, training behavior, and label quality.

Model outputs were treated as behavior events, then connected to dashboard workspaces for customer behavior analysis, sales analytics, inventory control, and AI-assisted reporting.

• Converted raw detections into structured retail behavior events.
• Linked behavior signals with sales, inventory, and decision-support logic.
• Designed dashboard workspaces for live vision, heatmaps, revenue metrics, stock status, and AI summaries.
• Focused on turning labels such as touch, hold, and item removed into business meaning.

RIOS also explored lightweight camera input using ESP32-S3-CAM as an early step toward edge and IoT deployment.

• Explored ESP32-S3-CAM as a lightweight camera source for retail shelf monitoring.
• Used the hardware exploration to think about future deployment beyond laptop-based testing.
• Positioned edge input as a future direction for lower-cost store-level monitoring.

The stack combines computer vision, backend APIs, dashboard UI, database storage, analytics workflows, and edge-camera exploration.

The hardest part was not only building separate layers, but making computer vision, retail data, dashboards, and AI reporting feel coherent as one decision system.

• Touch and hold actions can be visually similar, making behavior labeling and evaluation difficult.
• Limited dataset size, class imbalance, occlusion, and camera angle consistency affected model generalization.
• Computer vision labels are not valuable by themselves; they become useful when connected to business outcomes.
• Integration is not only about connectivity. It is about interpretability.
• The model is not the product. It is the signal generator for the system.

The next version would focus on stronger live inference, better signal quality, and more operational workflows for real retail environments.

• Connect live CV inference directly to streamed dashboard updates.
• Strengthen correlation between shelf behavior and downstream transactions.
• Add alerts for low conversion, stock pressure, and placement anomalies.
• Support richer store maps, shelf zones, and product-level event tracking.
• Improve edge-device and IoT deployment readiness.
• Add role-based views for store managers, merchandisers, and inventory teams.

RIOS showed me that detecting behavior is easier than making it meaningful. The real value comes from designing a system where behavior signals can be interpreted alongside sales, inventory, and operational KPIs.

• The strongest direction was treating RIOS as a decision system, not just a model or dashboard.
• The project improved my understanding of behavior -> insight -> action system design.
• The main lesson: a model becomes more useful when its output is connected to real workflows.