Shelf-Mounted Camera System for Accurate Cashier-Less Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automated shopping experiences are inaccurate and costly due to limitations in camera and sensor technology, requiring extensive store remodels and maintenance, and are not eco-system agnostic, making them difficult to integrate into existing retail environments.
Innovation Solution
A pluggable camera system equipped with AI and embedded computer vision algorithms mounted on shelf units to identify shoppers and track their activities, providing real-time inventory management and shopper behavior analytics without the need for extensive store redesign or external calibration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional camera and sensor systems are used for automated shopping, then the system can detect shopper activities, but the accuracy is insufficient due to technology limitations and customer attempts to fool the system
Solution Approach 1:
The patent combines multiple detection modalities (weight sensors, computer vision cameras, RFID readers) into an integrated detection system. The weight sensors on shelves detect item removal and placement, while computer vision algorithms analyze video feeds to identify shopper actions and verify item interactions. RFID tags on products provide additional verification. This multi-sensor fusion approach cross-validates detections, making it difficult for customers to fool the system and significantly improving detection accuracy compared to single-modality systems.
Solution Approach 2:
The patent introduces AI-powered computer vision algorithms as an intermediary layer between the physical shopping environment and the detection system. These algorithms process raw video data, identify shopper behaviors, and correlate visual events with sensor data. The AI intermediary interprets complex scenes, distinguishes between legitimate shopping actions and attempted system circumvention, and provides contextual understanding that raw sensor data alone cannot deliver, thereby enhancing both accuracy and reliability.
2Extent of automation
If extensive camera systems are installed for automated shopping detection, then the system can monitor shopper behavior, but the installation and maintenance costs increase substantially
Solution Approach 1:
The patent segments the automated detection system into modular, distributed components: individual shelf units with integrated weight sensors, standalone computer vision cameras positioned at strategic locations, and distributed processing nodes. Each shelf or zone operates semi-independently, allowing incremental deployment where systems can be installed shelf-by-shelf or zone-by-zone rather than requiring complete store overhaul. This segmentation reduces upfront installation costs and allows phased implementation.
Solution Approach 2:
The system employs AI algorithms that automatically calibrate and adjust detection parameters without requiring manual intervention. Computer vision systems self-calibrate by learning from observed shopper behaviors and item placements. Weight sensor thresholds are automatically adjusted based on historical data. This self-service capability eliminates the need for costly professional installation and ongoing maintenance by technical experts, allowing store staff to manage the system with minimal training.
3Extent of automation
If camera systems require store remodel or layout re-construction for installation, then the automated shopping experience can be implemented, but the disruption to store operations increases
Solution Approach 1:
The system is designed as modular units that can be attached to existing shelves and fixtures without structural modifications. Computer vision cameras mount to existing ceiling structures or shelving, and weight sensors attach to shelf edges using non-invasive methods. This segmentation allows installation in individual zones without requiring complete store closure or layout reconfiguration, enabling operations to continue in unaffected areas during installation.
Solution Approach 2:
The patent implements virtual calibration and pre-positioning of detection zones using software configuration before physical installation. Detection parameters, camera angles, and sensor thresholds are pre-configured based on digital store models, allowing rapid deployment once hardware is in place. This preliminary software preparation minimizes on-site adjustment time and reduces the duration of operational disruption during installation.
4Measurement precision
If RFID tags or sensors are embedded in shopping baskets and carts, then item tracking can be improved, but the system complexity and maintenance requirements increase
Solution Approach 1:
The patent extracts the tracking functionality from the shopping baskets and carts themselves and relocates it to fixed infrastructure elements. Instead of embedding sensors in mobile containers, the system uses fixed weight sensors on shelves to detect when items are removed and placed in baskets, and fixed RFID readers at strategic points to detect tagged items. This extraction eliminates the complexity of powering and maintaining sensors in mobile devices while achieving equivalent or superior tracking accuracy through fixed detection points.
Data Source
AI summary
Disclosed is a method of predicting shopping events. The method includes obtaining a first set of events from a sensor and a second set of events from a camera, wherein the camera captures one or more users in front of a shelf unit; determining whether a first timestamp from the obtained first set of events and a second timestamp from the obtained second set of events are within a same time interval; determining whether a first bin number matches a second bin number based on a determination that the first timestamp and the second timestamp are within the same time interval; and generating a shopping event for a user based on a determination that the first bin number matches the second bin number, wherein the user is associated with at least one of the obtained first set of events and the second set of events.