Automated LiDAR Road Reference Object Annotation for ADS Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automated driving systems face challenges in efficiently generating annotated training data for perception algorithms, particularly in handling complex traffic situations, due to the reliance on manual human intervention which can lead to inconsistencies and inefficiencies.
Innovation Solution
A method utilizing a Road Reference Object (RRO) prediction neural network to automatically generate annotated training data by predicting RRO positions in frames from LiDAR sensors, and matching these across multiple frames to form a global RRO position data set, reducing the need for manual intervention.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual human operators are used to produce annotated training data, then the training data can be generated with human judgment and correction, but the process becomes time-consuming and inconsistent
Solution Approach 1:
The system performs preliminary automated annotation using the trained perception algorithm before human operators review the data. This pre-annotation step prepares the data in advance, reducing the time human operators need to spend on each frame while maintaining high accuracy through subsequent human verification of the pre-processed annotations.
2Adaptability or versatility
If more complex models are used to handle increased variety of traffic situations, then the automated driving system can handle more scenarios, but the training data requirements increase and manual annotation becomes more difficult
Solution Approach 1:
The annotation process is segmented into automated processing using the perception algorithm and human review for complex cases. This segmentation allows the system to handle diverse traffic situations with automated tools while reserving human expertise for edge cases, managing the complexity of training diverse models without proportionally increasing manual annotation burden.
3Reliability
If human operators manually identify objects in sensor data, then the training data can be accurately annotated, but the productivity decreases and consistency varies between operators
Solution Approach 1:
The perception algorithm serves itself by generating annotations that can be directly used for training, reducing dependency on continuous human intervention. The system automatically processes sensor data, generates annotations, and makes them available for training, achieving both consistency through automated processing and high productivity through scalable algorithmic operation.
4Quantity of substance
If vast amounts of training data are required to train adequate perception models, then the model performance improves, but the bottleneck in manual data production limits further improvements
Solution Approach 1:
The manual mechanical process of human operators annotating data is replaced with an automated electronic system using the perception algorithm. This substitution enables massive scaling of training data production, as the algorithm can process sensor data at much higher speeds and volumes compared to human operators, removing the productivity bottleneck while maintaining or improving annotation quality.
Data Source
AI summary
A method for generating an annotated training data set for training a perception algorithm of an ADS of a vehicle is disclosed. The method includes obtaining a sequence of frames captured by a LiDAR sensors, predicting, using a road reference object (RRO) prediction neural network, an RRO position data set for each of a sub-set of the frames, wherein each RRO position data set includes RRO position data sub-sets for one or more RROs. Each RRO position data sub-set is related to spatial information of one RRO found in the frames, matching the PRO position data sub-sets of one frame with the PRO position data sub-sets of other frame to populate a global RRO position data set, wherein the global RRO position data set includes global RRO position data sub-sets, and forming the annotated training data set based on the sequence and the global RRO position data set.