Uncertainty-Based LIDAR Data Mining for Autonomous Vehicle Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous driving vehicles face challenges in training machine learning models for point cloud analysis due to the unstructured and large nature of point cloud data, limited labeled datasets, and resource-intensive processing requirements, which hinder efficient object detection and classification.
Innovation Solution
A method is implemented where a machine learning model onboard an autonomous driving vehicle computes confidence values for LIDAR data frames and uploads only those frames with obstacle detection below a confidence threshold to offboard storage, optimizing processing and bandwidth usage by selectively uploading data within a critical zone and adjusting thresholds based on data volume.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If all LIDAR data frames are uploaded for training, then training data sufficiency is improved, but bandwidth consumption and processing requirements increase
Solution Approach 1:
The patent extracts only the most valuable LIDAR data frames for training by using uncertainty estimation to identify frames with low confidence predictions. Instead of transferring all data, the system selectively extracts and uploads only those frames that are most needed for improving model accuracy, thereby reducing bandwidth consumption while maintaining training effectiveness.
Solution Approach 2:
The patent changes the parameter of data selection criteria from random or complete data transfer to uncertainty-based selective transfer. By monitoring model confidence scores and uploading only frames below a confidence threshold, the system transforms the data transfer process into an intelligent, parameter-driven selection mechanism that optimizes bandwidth usage.
2Measurement precision
If machine learning models process all point cloud data, then detection accuracy is improved, but computational resources and processing time increase
Solution Approach 1:
The patent applies partial action by processing only the subset of LIDAR frames that require attention according to uncertainty estimates. Instead of computationally intensive processing of all frames, the system identifies and processes only those frames with low confidence predictions, reducing overall computational power requirements while maintaining detection accuracy for critical objects.
Solution Approach 2:
The patent implements feedback through continuous monitoring of model confidence scores and using this information to guide future processing decisions. The system learns from previous predictions and adjusts which frames require processing, creating a feedback loop that optimizes computational resource allocation based on actual model performance and data quality.
3Reliability
If uncertainty threshold is set low, then data mining effectiveness is improved, but number of uploaded frames increases
Solution Approach 1:
The patent makes the uncertainty threshold dynamic rather than fixed, allowing it to adapt based on model performance, data distribution, and training needs. The threshold can be adjusted in real-time to balance the trade-off between training data quality and upload volume, optimizing the system for different operational conditions without requiring manual reconfiguration.
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
This approach reduces onboard processing requirements and bandwidth usage while improving the training of machine learning models for object detection and classification, enhancing the operational efficiency of autonomous vehicles.
Implementation Method 1
Light Detection and Ranging (LIDAR) systems in autonomous vehicles use a remote sensing technology that measures distances by illuminating targets with laser light and analyzing the reflected signals
Implementation Method 2
Light Detection and Ranging (LIDAR) systems in autonomous vehicles use a remote sensing technology that measures distances by illuminating targets with laser light and analyzing the reflected signals
Data Source
AI summary
The present disclosure provides a system and method that analyzes, onboard an autonomous driving vehicle (ADV), a frame of LIDAR data to identify one or more obstacles in the frame of LIDAR data. The system and method compute, by a machine learning model onboard the ADV, a confidence value for each of the one or more obstacles to produce one or more confidence values, wherein the one or more confidence values indicate a level of prediction certainty of the machine learning model. The system and method determine whether at least one of the one or more confidence values is below a confidence threshold. The system and method upload the frame of LIDAR data from the ADV to an offboard storage area based on determining that at least one of the one or more confidence values is below the confidence threshold.