3D Object Recognition via 2D Contour Projection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computer vision systems for object recognition and localization in 3D environments face limitations such as low processing speeds, poor generality, and the need for remote cameras, making them unsuitable for recognizing and classifying larger objects or those in complex orientations, which is a bottleneck in automation, recycling, and navigation applications.
Innovation Solution
A method that identifies primitives in digital images through gradient generation, curvature determination, and feature extraction, allowing for the classification and localization of 3D objects using a database of numerical descriptors for training images, enabling recognition and localization of objects regardless of camera position or object size, and allowing for partial occlusion and poor image segmentation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If standard computer vision systems are used for object recognition, then processing speed is improved, but the system requires remote camera positioning and cannot handle objects larger than 5-10 cm
Solution Approach 1:
The patent changes the fundamental parameters of the recognition model by incorporating perspective projection geometry and 3D spatial relationships. This allows the system to process objects at various distances and sizes by mathematically transforming 2D image coordinates into 3D object space, eliminating the need for fixed remote camera positioning while maintaining high processing speeds.
Solution Approach 2:
The system transitions from 2D image space analysis to 3D object space reconstruction by introducing depth information through perspective projection models. This dimensional transformation enables the camera to be positioned closer to objects and handles various object sizes by mapping 2D contours onto 3D reference models, resolving the limitation of fixed camera positioning.
2Productivity
If computer vision systems are designed for high processing speed, then generality is improved, but the systems cannot guide robots to grasp randomly oriented parts piled in boxes and pallets
Solution Approach 1:
The patent creates a universal recognition model that simultaneously handles multiple object orientations and positions by using 3D reference models that can be transformed to match any viewpoint. The system processes random orientations efficiently by comparing 2D image contours with projected 3D model contours from various angles, achieving both high speed and versatility in guiding robotic grasping.
Solution Approach 2:
The system performs preliminary 3D model matching and contour comparison to quickly eliminate impossible orientations before detailed analysis. By pre-computing projection relationships and using hierarchical matching strategies, the system rapidly determines feasible grasping configurations for randomly oriented parts without sacrificing processing speed.
3Device complexity
If 2D matching methods are used for object recognition, then implementation is simplified, but recognition accuracy for 3D objects is reduced
Solution Approach 1:
The patent introduces 3D reference models as an intermediary between 2D image data and object recognition. The system projects 3D model contours onto 2D image planes and compares them with actual image contours, using the 3D model as a mediator to bridge the gap between simplified 2D processing and accurate 3D localization, achieving both computational efficiency and precision.
Data Source
AI summary
The invention relates to a system and method for recognition, classification and spatial localization of bounded 3D-objects. In particular it relates to a computerized method for recognition, classification and localization of objects. The method comprises generation of a training database based on a large number of training views recorded by a camera or constructed using a CAD representation of an object. Characteristic curves are derived from the training views, and primitives of the curves are detected. Intrinsic and extrinsic descriptors of features are stored in the database together with data about the object class and pose of the view. Finally the recognition takes place in two stages: First the intrinsic descriptors of the recognition view are compared with those of the database. Second, among the best matching features it is explored which features agree mutually in the sense that they suggest the same object class at the same pose.