Robot Target Recognition Using LiDAR Projection and Image Feature Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in accurately identifying a target object using a combination of camera and LiDAR data for robot navigation, particularly in processing and integrating 2D image data with 3D spatial data to enhance path planning and object recognition.
Innovation Solution
A robot control apparatus and method utilizing LiDAR and a camera to project point clouds onto a designated surface, process images through a neural network model, and employ a classifier group with Gaussian probability distributions to identify target objects by analyzing feature maps and pixel values, enabling real-time target recognition and path planning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If camera and LiDAR data are integrated for target identification, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The patent introduces an intermediary processing system that receives data from both camera and LiDAR sensors, processes them through a unified algorithm, and produces integrated target identification results. This intermediary layer manages the complexity of sensor fusion while maintaining high measurement precision through coordinated processing of visual and spatial data.
Solution Approach 2:
The patent merges camera image data and LiDAR point cloud data into a unified target identification process. By combining the visual recognition capabilities of the camera with the spatial measurement capabilities of LiDAR, the system achieves improved measurement precision while managing device complexity through integrated processing algorithms.
2Measurement precision
If point cloud projection and neural network processing are used, then target identification accuracy is improved, but computing time increases
Solution Approach 1:
The patent performs preliminary projection of the point cloud onto a 2D surface before neural network processing. This preliminary action simplifies the data structure and reduces the computational complexity of subsequent neural network operations, thereby reducing processing time while maintaining target identification accuracy through pre-organized spatial information.
Solution Approach 2:
The patent segments the processing pipeline into distinct stages: point cloud acquisition, projection onto 2D surface, feature extraction, and neural network classification. This segmentation allows each stage to be optimized independently, reducing overall computing time while maintaining high target identification accuracy through specialized processing at each stage.
3Measurement precision
If classifier group with multiple Gaussian distributions is employed, then classification accuracy is improved, but device complexity increases
Solution Approach 1:
The patent employs a classifier group that utilizes multiple Gaussian probability distributions with different parameters (means, variances) to model various target characteristics. By changing and adjusting these parameters based on training data, the system achieves high classification accuracy while managing algorithmic complexity through parameter optimization rather than structural complexity.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Enables accurate and efficient identification of target objects, facilitating effective path planning for robots while maintaining low operational costs, and providing reliable navigation capabilities.
Implementation Method 1
based on obtaining the point cloud by use of the LiDAR
Implementation Method 2
inputting a portion of an image obtained by use of the camera, the image including a visual object corresponding to the virtual object
Data Source
AI summary
A robot control apparatus can include light detection and ranging (LiDAR), a camera, a memory storing a classifier group including a plurality of classifiers and a neural network model, and a processor. The processor can be configured to project a point cloud corresponding to an external object onto a designated surface to obtain a virtual object represented in two dimensions, based on obtaining the point cloud, input a portion of an image obtained by use of the camera, which includes a visual object corresponding to the virtual object, to the neural network model, based on identifying the visual object in the image, and input a designated number of feature maps for the portion of the image to the classifier group to identify whether the external object corresponding to the visual object is a target object, based on obtaining the feature maps.