LIDAR Fault Detection Using Synthetic Point Cloud Abnormal Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting LIDAR malfunctions are limited to specific environments and cannot detect malfunctions when the LIDAR is not operating.
Innovation Solution
A method using artificial intelligence technology to generate a LIDAR-abnormal dataset based on a normal dataset, train a LIDAR malfunction detection model, and use it to detect malfunctions in various environments, including a pre-trained model that determines faults based on point cloud data analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing methods use test patterns or targets in specific environments to detect LIDAR malfunctions, then detection can be performed in controlled conditions, but detection capability is limited to specific environments only
Solution Approach 1:
The patent creates synthetic LIDAR abnormal datasets by copying and transforming normal LIDAR point cloud data through various simulated malfunction scenarios. This allows the model to learn from artificial copies of abnormal conditions without requiring actual physical test targets or controlled environments, thereby enabling detection across diverse real-world environments.
Solution Approach 2:
The patent applies parameter changes by systematically modifying point cloud data parameters (such as point density, spatial distribution, intensity values) to simulate different LIDAR malfunction states. This transforms normal operational data into abnormal training samples, enabling the model to recognize various failure modes across different environmental conditions.
2Reliability
If existing methods assume LIDAR operation to detect malfunctions, then detection can be performed during normal operation, but malfunctions cannot be detected when LIDAR does not operate
Solution Approach 1:
The patent performs preliminary action by training the detection model in advance using both normal and synthetically generated abnormal datasets before actual LIDAR deployment. This pre-training enables the model to detect malfunctions whether the LIDAR is operating or not, as the model has already learned to recognize abnormal patterns during the training phase.
Solution Approach 2:
The patent generates synthetic abnormal datasets by copying and transforming normal LIDAR data to create virtual malfunction scenarios. This allows the model to learn failure detection without requiring actual LIDAR failures or non-operational states during training, thereby enabling comprehensive detection coverage across all operational states.
3Measurement precision
If actual LIDAR abnormal data is collected for model training, then training data reflects real malfunction conditions, but data collection costs and time increase significantly
Solution Approach 1:
The patent creates synthetic abnormal training data by copying and transforming existing normal LIDAR point cloud data through simulated malfunction operations. This eliminates the need for time-consuming collection of actual abnormal LIDAR data from real failure scenarios, while still providing sufficient training examples for the model.
Solution Approach 2:
The patent uses computationally inexpensive synthetic data generation methods instead of expensive and time-consuming real-world data collection. The synthetic datasets can be rapidly generated through algorithmic transformations of normal data, providing a cost-effective alternative to physical data collection campaigns.
Data Source
AI summary
A method and an apparatus for detecting malfunction of LIDAR, and a method and an apparatus for generating data therefor are disclosed.According to an aspect of the present disclosure, there is provided a method for detecting malfunction of a LIDAR, including: acquiring point cloud data from the LIDAR; determining whether the number of points included in the point cloud data is less than a point count threshold; and determining whether the LIDAR has a fault based on the point cloud data by using a LIDAR malfunction detection model when the number of points is greater than or equal to the point count threshold, wherein the LIDAR malfunction detection model is a pre-trained model based on a normal dataset and at least one LIDAR-abnormal dataset, the normal dataset includes preset statistics and point cloud data acquired from the LIDAR in a normal state, and the LIDAR-abnormal dataset includes point cloud data generated based on the normal dataset.