3D Object Position Detection Using Simple-Shape Ensemble Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection techniques, such as template matching and SSD, face difficulties in detecting objects with shapes different from those envisioned in advance, particularly in environments where teaching is required for picking robots, and lack sufficient learning data in the field of picking work.
Innovation Solution
A position detection program utilizing an ensemble learning model based on multiple neural networks trained on datasets of simple shapes, which combines the reliability outputs from each model to detect objects with a shape different from the anticipated, thereby reducing the need for extensive teaching and enhancing detection versatility.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If template matching or SSD is used for object detection, then detection speed and simplicity are improved, but the ability to detect objects with shapes different from those envisioned in advance deteriorates
Solution Approach 1:
The invention segments the object detection task into multiple specialized neural networks, each trained to detect specific simple shapes (spheres, cubes, cylinders, etc.). Instead of using a single general-purpose detector, the system divides the detection problem into multiple specialized sub-problems, with each neural network segment handling a particular shape category.
Solution Approach 2:
The invention merges the outputs of multiple specialized neural networks by combining their reliability scores. The system integrates the detection results from each shape-specific network through a combination process that aggregates reliability information, allowing the system to leverage the strengths of multiple specialized detectors while maintaining high detection speed.
2Measurement precision
If teaching is required for picking robots to detect objects, then detection accuracy for known objects is improved, but the time and effort required for teaching increases
Solution Approach 1:
The invention performs preliminary action by pre-training multiple neural networks on various simple shapes before actual object detection. These pre-trained networks are stored in the system and can be directly applied to detect objects during operation, eliminating the need for time-consuming teaching processes for each new object while maintaining high detection accuracy.
Solution Approach 2:
The system achieves universality by creating neural networks that can detect multiple types of simple shapes (spheres, cubes, cylinders, cones, etc.) using the same detection framework. This multi-functional approach allows a single system to handle various object shapes without requiring separate teaching for each shape type, significantly reducing teaching time while maintaining accuracy.
3Adaptability or versatility
If deep learning is used to detect objects with unknown shapes, then detection versatility is improved, but the amount of learning data required increases
Solution Approach 1:
The invention applies local quality by training each neural network on specific local characteristics of simple shapes rather than requiring comprehensive data on all possible object variations. Each network focuses on detecting particular geometric features (sphericity, cubicity, cylindricality) with high precision, allowing versatile detection with limited specialized training data for each shape type.
Data Source
AI summary
A non-transitory computer-readable storage medium storing a position detection program which causes a processor to perform processing for object recognition, the processing includes: acquiring a plurality of pieces of three-dimensional data of simple shapes that are not similar to each other; carrying out learning by using the plurality of acquired pieces of data; acquiring an image obtained by imaging by an imaging unit; and detecting a position of an object from the acquired image by using a first learning model generated based on the learning.