3D Carton Pile Vision Detection Without Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automated systems for single box picking and decanting require pre-known information about box arrangements, are prone to errors due to arbitrarily selected parameters, and necessitate extensive training and retraining for new objects, making them inefficient and costly.
Innovation Solution
A machine vision-based system using 3-D sensors and probabilistic image processing to identify and rank hypotheses about box locations and orientations without training, eliminating the need for feature extraction and parameter selection, and allowing for rapid adaptation to new objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If pre-defined parameters and training data are used for automated box picking, then system setup is simplified, but adaptability to new box types and arrangements deteriorates
Solution Approach 1:
The system performs self-training by automatically learning box characteristics and arrangements from images during operation, without requiring external training data or parameter configuration. The machine learning model updates itself continuously to adapt to new box types and pile configurations autonomously
Solution Approach 2:
The system performs preliminary learning during initial operation phases by analyzing images of boxes and their arrangements, building a knowledge base that enables rapid adaptation to new box types and configurations before actual picking operations begin
2Speed
If classical image processing with feature extraction is used, then processing speed is improved, but measurement precision deteriorates due to arbitrarily selected parameters
Solution Approach 1:
The patent replaces classical image processing mechanics (feature extraction, parameter selection) with a machine learning-based approach that automatically learns optimal features and parameters from data, eliminating arbitrary parameter selection and improving measurement precision while maintaining processing speed
3Measurement precision
If extensive training and retraining is performed for new objects, then measurement precision is improved, but loss of time increases
Solution Approach 1:
The system performs self-training by automatically learning box characteristics and arrangements from images during operation, without requiring external training data or parameter configuration. The machine learning model updates itself continuously to adapt to new box types and pile configurations autonomously
Solution Approach 2:
The learning process continues continuously during normal operation rather than requiring separate training phases, allowing the system to maintain high measurement precision while minimizing time loss through ongoing incremental learning from actual operational data
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Facilitates fast, reliable, and cost-effective unloading of cartons by eliminating training requirements and reducing errors, enabling efficient handling of varied box types and orientations.
Implementation Method 1
The at least one sensor has a set of radiation sensing elements which detect reflected, projected radiation to obtain 3-D sensor data
Data Source
Figure 1~2
Figure 3~4
Figure 5~7
AI summary
A machine vision-based method and system to facilitate the unloading of a pile of cartons within a work cell are provided. The method includes the step of providing at least one 3-D or depth sensor having a field of view at the work cell. Each sensor has a set of radiation sensing elements which detect reflected, projected radiation to obtain 3-D sensor data. The 3-D sensor data including a plurality of pixels. For each possible pixel location and each possible carton orientation, the method includes generating a hypothesis that a carton with a known structure appears at that pixel location with that container orientation to obtain a plurality of hypotheses. The method further includes ranking the plurality of hypotheses. The step of ranking includes calculating a surprisal for each of the hypotheses to obtain a plurality of surprisals. The step of ranking is based on the surprisals of the hypotheses.