ICP Alignment Verification for Reliable Autonomous Vehicle Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately tracking objects using the iterative closest point (ICP) algorithm, as incorrect alignments can impact object tracking, behavior prediction, and trajectory planning, necessitating a method to verify the accuracy of ICP alignments.
Innovation Solution
A machine-learned model is trained using human-labeled LIDAR sensor data to determine the trustworthiness of ICP alignments, allowing autonomous vehicles to disregard or adjust data based on alignment accuracy, and potentially utilize other sensors like radar for velocity determination.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the ICP algorithm is used to determine correspondences between LIDAR point clouds, then object tracking and velocity determination are enabled, but alignment accuracy deteriorates due to incorrect alignments
Solution Approach 1:
A machine-learned validation model is introduced as an intermediary between the ICP algorithm and the object tracking system. This model validates ICP alignment results by comparing them against ground truth data, filtering out incorrect alignments before they affect tracking accuracy. The validation model acts as a mediator that ensures only reliable alignment data is used for velocity determination and trajectory planning.
2Reliability
If ICP alignment data is used for velocity determination, then object behavior prediction is improved, but reliability deteriorates when alignments are incorrect
Solution Approach 1:
A feedback mechanism is implemented where the validation model continuously assesses the quality of ICP alignments and provides reliability scores. When alignments are deemed unreliable, the system adjusts its behavior prediction accordingly, either by disregarding the problematic alignment data or by increasing uncertainty in predictions. This feedback loop ensures that behavior prediction maintains high reliability even when ICP alignments are occasionally incorrect.
3Measurement precision
If the validation model is trained with human-labeled data, then alignment verification accuracy is improved, but system complexity increases
Solution Approach 1:
The validation model is trained in advance using human-labeled ground truth data before deployment in the autonomous vehicle system. This preliminary training phase allows the model to learn the characteristics of correct versus incorrect alignments offline, reducing the computational burden and complexity during real-time operation. The training data is collected and processed beforehand, so that during actual driving, the model can quickly validate alignments without adding significant system complexity.
Data Source
AI summary
Aspects of the disclosure relate to training and using a model for verifying accuracy of ICP alignments or alignments between data points using an iterative closest point algorithm. For instance, a model may be trained using ICP alignment data, including alignments between an object appearing in LIDAR sensor frames. The training may also include setting a definition for a trusted ICP alignment. In this regard, the model may be trained such that, n response to receiving additional LIDAR sensor frames and corresponding additional ICP alignment data, output a value indicative of whether the additional ICP alignment data is trusted according to the definition. The model may then be used to control a vehicle in an autonomous driving mode by determining whether alignment data for object determined using the ICP algorithm should be trusted.