Autonomous Shelf Product Recognition Using Template Vector Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing stock tracking systems struggle to efficiently and accurately identify and recognize products on inventory structures within a store, particularly in dynamic retail environments with varying lighting conditions and product orientations.
Innovation Solution
A method utilizing a mobile robotic system to capture images of inventory structures, extract visual features, and populate a multi-dimensional space with template vectors labeled with product identifiers, enabling autonomous product detection and recognition by calculating similarity scores between new images and template vectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional stock tracking systems are used to identify products on inventory structures, then the system can perform basic product recognition, but the accuracy is insufficient in dynamic retail environments with varying lighting conditions and product orientations
Solution Approach 1:
The patent transforms product identification from traditional 2D image recognition to 3D point cloud analysis. The robotic system captures spatial data in three dimensions, creating point clouds that represent product geometry and position. This dimensional transformation enables accurate product identification regardless of lighting conditions or orientation, as the 3D structural features remain consistent unlike 2D visual appearances which vary with environmental factors.
Solution Approach 2:
The patent replaces traditional optical/image-based recognition systems with a robotic system equipped with depth sensors and point cloud processing capabilities. This substitution from optical to spatial-mechanical sensing allows the system to identify products based on their geometric structure rather than visual appearance, thereby achieving accuracy independent of lighting conditions and product orientation.
2Productivity
If manual product identification methods are used, then the system can maintain inventory records, but the process is time-consuming and reduces productivity
Solution Approach 1:
The robotic system autonomously performs product identification without human intervention. It independently navigates inventory structures, captures point cloud data, processes the spatial information, and updates inventory records automatically. This self-service capability eliminates the time loss associated with manual product identification while maintaining high productivity in inventory tracking operations.
Solution Approach 2:
The robotic system continuously performs inventory tracking operations without interruption. It systematically moves through inventory structures, continuously capturing and processing point cloud data to identify products and update records in real-time. This continuous operation maximizes productivity while minimizing the time loss that would occur with periodic or manual inventory checks.
3Adaptability or versatility
If existing image-based recognition systems are used, then the system can identify products, but it struggles with varying product orientations and lighting conditions
Solution Approach 1:
The patent transitions from 2D image-based recognition to 3D point cloud analysis. By representing products in three-dimensional space, the system can accurately identify products regardless of their orientation. The point cloud data captures the geometric structure of products from multiple angles simultaneously, allowing the system to recognize products in any orientation without the accuracy degradation that plagues 2D image-based systems.
Solution Approach 2:
The robotic system dynamically adapts its data capture and processing approach based on the spatial configuration of products. The point cloud processing algorithms dynamically adjust to accommodate varying product orientations and positions on inventory structures, maintaining high recognition accuracy across diverse product arrangements and orientations that would challenge static image-based systems.
Data Source
AI summary
One variation of a method includes: accessing an image of an inventory structure captured by a robotic system navigating within a facility; detecting an object occupying a slot in the inventory structure depicted in the image; extracting a set of visual features from the image; representing the set of visual features in a vector; projecting the vector into a multi-dimensional space populated template vectors representing product units of verified product types within the facility; calculating a similarity score between the set of visual features and template visual features represented in a cluster of template vectors in the multi-dimensional space, based on proximity between the vector and the cluster of template vectors; and, in response to the similarity score exceeding a threshold score, identifying the object as a product unit of a first product type affiliated with a first product identifier associated with the cluster of template vectors.