Robot Object Detection Using Point Pair Feature Voting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for sensor-based object recognition in robot and autonomous driverless transport systems are complex and lack robustness, particularly in decision-making processes, leading to inefficient object detection and control.
Innovation Solution
A method using 2D surface contour models with point pair features and a voting grid for pose hypothesis determination, combined with consistency checks, to simplify and enhance the reliability of object recognition, allowing for efficient control of robots and autonomous systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If three-dimensional point cloud processing with multiple point pair features is used for object recognition, then object detection capability is improved, but computational complexity and processing time increase significantly
Solution Approach 1:
The patent extracts only the essential geometric parameters (distance between points and angles of normals) needed for object recognition, forming simplified point pair features. This extraction approach maintains object detection capability while significantly reducing computational complexity by focusing only on the most relevant geometric characteristics rather than processing all possible point cloud attributes.
Solution Approach 2:
The patent segments the complex object recognition task into distinct stages: generating point pair features from the point cloud, voting for pose hypotheses based on these features, and finally accepting or rejecting hypotheses through consistency checks. This segmentation allows each stage to be optimized independently, reducing overall computational burden while maintaining detection accuracy.
2Ease of operation
If a voting and clustering method is used for final detection decision, then processing simplicity is improved, but detection reliability and robustness deteriorate
Solution Approach 1:
The patent implements a feedback mechanism where pose hypotheses are subjected to consistency checks that verify whether the hypothesized object pose is compatible with the actual sensor measurements. This feedback loop ensures that only physically plausible hypotheses are accepted, significantly improving detection reliability while maintaining the simplicity of the voting approach.
Solution Approach 2:
The patent performs preliminary consistency checks on pose hypotheses before final acceptance. By pre-filtering hypotheses based on geometric and physical constraints before the final decision, the system ensures robust detection while keeping the overall process simple and efficient.
3Measurement precision
If comprehensive point pair feature analysis is performed, then object recognition accuracy is improved, but processing time and computational resources increase
Solution Approach 1:
The patent uses a partial action approach by forming point pair features from a subset of point cloud points rather than all possible combinations. This selective feature formation maintains sufficient recognition accuracy while dramatically reducing the number of computations required, thus decreasing processing time.
Solution Approach 2:
The patent performs preliminary formation of point pair features and voting for pose hypotheses before conducting detailed consistency checks. This preliminary processing allows the system to quickly eliminate obviously incorrect hypotheses, reducing the time needed for comprehensive analysis while maintaining accuracy.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
The present invention relates to a method for controlling a robot and/or an autonomous driverless transport system on the basis of sensor-based detection of objects. In this case, point pair features of the 2-D surface contours are produced on the basis of 2-D surface contours of the objects to be detected. A point cloud of the environment is captured by means of a distance sensor and a surface normal is estimated for each point, and corresponding point pair features of the environment are produced. In a voting method, environment features are compared with model features in order to efficiently generate pose hypotheses which are subjected to optimization and to a consistency check in further steps in order to finally be accepted or rejected as detection.