Autonomous Vehicle Perception Validation Using Consensus Modules
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately identifying perception errors, which can lead to false negatives or false positives in navigation, due to the reliance on a single perception module without independent validation, potentially causing safety issues and inefficiencies.
Innovation Solution
Implementing one or more independent perception validation modules, each with a machine-learning model trained on different data, to corroborate perception outputs and determine a ground-truth output through consensus rules, such as voting mechanisms, to identify and escalate errors for human review.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single perception module is used for autonomous vehicle navigation, then the device complexity is reduced, but the reliability of perception outputs deteriorates due to lack of independent validation
Solution Approach 1:
The perception system is segmented into multiple independent perception modules, each processing sensor data through different machine learning models. This segmentation allows independent validation of perception outputs without requiring a complete system redesign, thereby improving reliability while managing complexity through modular architecture
Solution Approach 2:
An intermediary validation layer is introduced that receives outputs from multiple perception modules and compares them to identify discrepancies. This intermediary component coordinates the validation process and determines ground-truth outputs through consensus rules, improving reliability while isolating the complexity of validation logic from individual perception modules
2Measurement precision
If multiple independent perception validation modules are implemented, then the accuracy of perception outputs is improved through error identification, but the device complexity increases due to additional modules and consensus mechanisms
Solution Approach 1:
Additional perception modules are implemented as copies of the primary perception module, each with identical or similar machine learning models. These copied modules produce redundant perception outputs that can be validated against each other, improving measurement precision while maintaining manageable complexity through standardized module design
Solution Approach 2:
The perception modules are designed with universal functionality to process various sensor data types and produce standardized perception outputs. This multi-functionality allows the same module architecture to serve both primary perception and validation purposes, reducing overall system complexity while improving accuracy through cross-validation
3Reliability
If perception outputs are validated through consensus rules and voting mechanisms, then the reliability of ground-truth determination is improved, but the processing time and productivity are reduced
Solution Approach 1:
The consensus mechanism is implemented with partial action by setting thresholds for validation. When perception outputs from multiple modules agree within acceptable margins, validation is completed quickly without requiring full consensus from all modules. This approach maintains high reliability for clear cases while improving productivity by avoiding excessive processing for unambiguous situations
Solution Approach 2:
The validation process uses periodic action by implementing hierarchical validation levels. Routine perception outputs undergo quick validation checks, while only disputed or uncertain outputs trigger more comprehensive consensus processes. This periodic intensification of validation maintains reliability for critical cases while improving overall processing speed through selective detailed validation
Data Source
AI summary
The disclosed technology provides solutions for validating/verifying perception outputs, e.g., from a perception module of an autonomous vehicle (AV) software stack. In some aspects, a process of the disclosed technology can include steps for providing sensor data to a perception module, receiving, from the perception module, a first perception output based on the sensor data, providing the sensor data to a validation module, and receiving, from the validation module, a second perception output based on the sensor data. In some aspects, the process can further include steps for determining if the first perception output corresponds with the second perception output. Systems and machine-readable media are also provided.