Sparse SLAM Object Tracking with Map Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Autonomous systems face challenges in accurately determining geometric information and 3D modeling of objects in unstructured environments, particularly in cluttered scenes where object recognition and segmentation are imperfect, leading to contamination between maps and limited usability in dynamic environments.

Innovation Solution

A system leveraging a sparse point-plane SLAM algorithm with a novel segment-classification algorithm manages multiple independent SLAM maps for independent 3D object detection and reconstruction, reducing CPU usage and power consumption by focusing on object manipulation rather than full environment reconstruction, and employing a computer-implementable method for object detection and manipulation using 3D sensors and manipulators.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a SLAM framework is used to simultaneously register static scene and track moving objects, then object tracking capability is improved, but map contamination risk increases when objects contact with environment

Engineering Contradiction:
Improveobject tracking accuracyVSAvoidmap contamination
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The system segments the scene into multiple independent SLAM maps, each dedicated to tracking a specific moving object. This segmentation prevents map contamination by isolating object features from static environment features, allowing each map to maintain purity while tracking its designated object independently.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system introduces scene segmentation as an intermediary mechanism between the SLAM framework and object tracking. The segmentation process acts as a filter that separates object-related features from static environment features before they are added to the SLAM map, preventing contamination while maintaining tracking accuracy.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Loss of information

If full environment reconstruction is performed, then complete scene understanding is achieved, but CPU usage and power consumption increase significantly

Engineering Contradiction:
Improvescene understanding completenessVSAvoidCPU usage and power consumption
Core Design Contradiction:
Loss of informationVSUse of energy by moving object

Solution Approach 1:

The system extracts and processes only the relevant moving objects from the full environment, rather than reconstructing the entire scene. By isolating and tracking only dynamic objects with dedicated SLAM maps, the system achieves sufficient scene understanding for manipulation tasks while dramatically reducing computational load and energy consumption.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system performs partial reconstruction focused specifically on moving objects rather than complete environment reconstruction. This selective approach applies SLAM processing only where necessary (on dynamic objects), achieving the minimum required scene understanding while minimizing CPU usage and power consumption.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3782119B1Detection, tracking and 3D modeling of objects with sparse RGB-d slam and interactive perception
Publication Date: 2024.07.03 MITSUBISHI ELECTRIC CORP
  • EP3782119B1 patent drawingFigure 1
  • EP3782119B1 patent drawingFigure 2
  • EP3782119B1 patent drawingFigure 3~4

AI summary

An object detection system for detecting and manipulating objects on a workspace includes a three dimensional (3D) sensor configured to acquire and transmit point clouds of a scene, each point cloud including one or more objects in the workspace, manipulator configured to move or grip each of the one or more objects, a memory to store the images and a computer executable program including an object detection-localization program, a segmentation program, a gripping-moving program and a geometry reconstruction program, a processor to perform the computer executable program using the images in connection with the 3D sensor, the memory and the manipulator. The computer executable program is configured to cause the processor to perform steps of generating frames including the point clouds and measurements, generating an initial map from a first frame, moving part of the one or more objects by using the manipulator using the first image, acquiring, after the moving step, a second frame, generating a first dynamic map and a static map from the initial map using the second frame, moving another part of the one or more objects based on the first dynamic map, splitting at least one dynamic map among the first dynamic map and the second dynamic maps when the at least one dynamic map includes a segment being outlier with respect to the one or more objects, reconstructing a 3D geometry of the one or more objects from the split map, and determining when the one or more objects are separated from each other.