Information Processing Apparatus for Image-Guided Low-Power 3D Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous mobile robots face challenges in efficiently and accurately clustering large amounts of three-dimensional point cloud data with limited processing power, necessitating high-performance processors that consume excessive power, which is not feasible for battery-driven robots or drones.
Innovation Solution
Estimate the presence area of objects and their number based on captured images, generate point cloud data from distance measurements, and perform clustering using machine learning to determine optimal starting points and cluster numbers, thereby narrowing down the point clouds to be processed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a high-performance processor is used to perform point cloud clustering in real time, then the processing speed and accuracy of clustering are improved, but the power consumption increases significantly
Solution Approach 1:
The patent segments the point cloud data processing by first performing 2D object detection on image data to identify object locations, then using these locations to guide 3D point cloud clustering. This division of processing stages allows the system to process only relevant point cloud regions rather than all points, reducing computational load and power consumption while maintaining real-time performance
Solution Approach 2:
The patent performs preliminary 2D object detection and location estimation on image data before processing the 3D point cloud data. This preliminary action identifies regions of interest, allowing the subsequent point cloud clustering to focus only on relevant areas, thereby reducing the overall processing time and power requirements while achieving real-time clustering performance
2Use of energy by moving object
If a processor for an embedded device is used to reduce weight, then the power consumption is reduced, but the processing performance of clustering is limited
Solution Approach 1:
The patent divides the clustering task into two segments: 2D object detection from images (performed on embedded device) and 3D point cloud clustering (performed on high-performance processor). This segmentation allows the embedded device to handle preliminary processing with low power consumption while the computationally intensive 3D clustering is offloaded to a high-performance processor, achieving both power efficiency and processing performance
Solution Approach 2:
The patent uses 2D object detection results as an intermediary to guide the 3D point cloud clustering process. The detected object locations and presence areas from the 2D image serve as a bridge, providing initial values and region constraints that significantly reduce the computational complexity of the subsequent 3D clustering, enabling embedded devices to achieve better processing performance with limited power consumption
3Ease of operation
If random selection of center point is used in clustering, then the implementation is simple, but the convergence speed of point cloud to cluster cloud is unstable
Solution Approach 1:
The patent performs preliminary 2D object detection to estimate object locations and presence areas before initiating 3D point cloud clustering. These preliminary results provide informed initial center points for the clustering algorithm, replacing random selection with data-driven choices that significantly improve convergence stability and speed while maintaining implementation feasibility
Solution Approach 2:
The patent uses feedback from 2D object detection results (object locations, presence areas) to guide and adjust the 3D point cloud clustering process. The detected 2D object information provides feedback that constrains and directs the clustering to focus on relevant regions, ensuring stable and reliable convergence while keeping the system relatively simple to implement
Data Source
AI summary
An information processing apparatus estimates a presence area of an object and the number of objects on the basis of a captured image, generates point cloud data from distance measurement information acquired by a distance measuring sensor, and recognizes the object by determining a point cloud, which is a target of clustering, in the point cloud data generated and the number of clusters on the basis of the presence area of the object and the number of the objects, which are estimated, and performing clustering on the point cloud data.