LiDAR Full Waveform Classification Using Dual Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current LiDAR data classification methods fail to fully utilize the additional information from full waveform signals, leading to inaccurate and time-consuming manual corrections, especially in classifying vegetation and urban areas, due to reliance on extracted characteristics rather than raw waveform data.
Innovation Solution
A classification apparatus utilizing two neural networks to process full waveform LiDAR data in its raw form, where the first neural network analyzes waveform development independently and the second network processes spatial coordinates and probability distributions to achieve precise classification without manual intervention or prior knowledge.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If extracted characteristics are used for classification, then processing speed is improved, but classification accuracy deteriorates due to loss of waveform information
Solution Approach 1:
The patent extracts specific features (maximum amplitude, width at maximum amplitude, full width at half maximum) from the complex waveform signal to create a simplified feature vector. This extraction process retains the most discriminative information while reducing data dimensionality, thereby maintaining classification accuracy while improving processing efficiency.
Solution Approach 2:
The patent introduces an intermediary feature extraction step that transforms raw waveform data into a compressed feature representation. This intermediary layer acts as a bridge between the detailed waveform information and the classification algorithm, preserving essential characteristics while reducing computational complexity.
2Measurement precision
If full waveform data is processed, then classification accuracy is improved, but processing time increases
Solution Approach 1:
The patent extracts specific features (maximum amplitude, width at maximum amplitude, full width at half maximum) from the complex waveform signal to create a simplified feature vector. This extraction process retains the most discriminative information while reducing data dimensionality, thereby maintaining classification accuracy while improving processing efficiency.
Solution Approach 2:
The patent applies partial action by selectively processing only the most informative portions of the waveform (peak region and half-maximum points) rather than analyzing the entire waveform continuously. This partial processing approach captures sufficient classification information while significantly reducing computational burden.
3Measurement precision
If manual corrections are applied to classification results, then accuracy is improved, but time consumption increases
Solution Approach 1:
The patent implements self-service by enabling the classification system to automatically correct its own errors through iterative refinement. The system uses feedback from initial classification results to adjust decision thresholds and re-classify ambiguous cases, achieving high accuracy without requiring manual intervention.
Solution Approach 2:
The patent incorporates feedback mechanisms where classification results are evaluated and used to refine subsequent classification decisions. The system adjusts classification parameters based on performance feedback, automatically improving accuracy through iterative optimization rather than manual correction.
4Measurement precision
If prior knowledge is required for classification, then accuracy is improved, but ease of operation deteriorates
Solution Approach 1:
The patent implements self-service by enabling the classification system to automatically correct its own errors through iterative refinement. The system uses feedback from initial classification results to adjust decision thresholds and re-classify ambiguous cases, achieving high accuracy without requiring manual intervention.
Solution Approach 2:
The patent automatically adapts classification parameters (such as decision thresholds and feature weights) based on the input data characteristics. This parameter adaptation occurs without requiring users to have prior knowledge or manually tune parameters, making the system both accurate and easy to operate.
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
The method enables accurate and efficient automatic classification of LiDAR data, achieving high precision and reducing processing time by directly exploiting all information from full waveform signals, as demonstrated by an accuracy of 93.06% in classifying terrain, vegetation, and urban features.
Implementation Method 1
laser source to illuminate the earth's surface and a photodiode to record the retro-diffused radiation
Implementation Method 2
The measurement of the signal return time provides the measurement of the distance between the instrument and the reflecting object
Implementation Method 3
When the object is not solid, or is not too dense (for example, tree branches), part of the laser beam can continue its trajectory and can be reflected by lower obstacles
Data Source
Figure 1~4
Figure 2
Figure 3
AI summary
Classification apparatus to classify full waveform data (S) from signals retro- reflected from points of detection of objects subjected to scanning by electromagnetic waves comprising a first neural network classification device (11) configured to receive at input and to process said full waveform data (S) and a second neural network data processing device (12), located and operatively connected downstream of the first classification device (11) and configured to supply at output the relative class of each object from which said full waveform data signal (S) has been reflected.