Overview
An ML-based perception model was developed for an autonomous vehicle to detect road and lane boundaries for both the vehicle lane and adjacent lanes. The model supported lateral localization, lane keeping, and situational awareness within the vehicle software architecture. 

The main vehicle-level hazards were (1) unintended lane departure and (2) collision with adjacent-lane traffic. The safety analysis focused on risk assessment by defining the model’s intended function, system interfaces, and linking vehicle-level hazards to failure modes throughout the model development pipeline. 

Challenges
The challenge was assessing the safety risk of the model within the overall autonomous vehicle perception system. This required identifying which model development pipeline issues could affect the safety of the vehicle and establishing a consistent and defensible methodology for evaluating those risks. 

Before this analysis could be performed, the model’s functions, responsibilities, and interfaces needed to be clearly identified and documented. This foundation was necessary to establish traceability between model development activities, perception failures, and vehicle-level hazards. 

Approach
A structured Machine Learning Failure Modes and Effects Analysis (ML-FMEA) was applied to improve traceability between perception failures and system-level hazards. The approach included: 

• Explicit definition of ML pipeline stages, including define scope, collect data requests, collect data, ingest data, validate data, construct ground truth or reward feedback, preprocess data, train model, tune model, analyze model, deploy model, validate model, and analyze model feedback 

• Mapping of perception failure modes to vehicle-level hazards such as unintended lane departure and adjacent-lane collision 

• Evaluation of safety-critical performance scenario variants, including rare lane geometries, temporary markings, and edge-case road layouts 

• Assessment of dataset coverage across weather conditions, geographic regions, and sensor configurations 

Additional attention was given to identifying hidden dataset issues such as annotation drift and long-term distribution shifts. 

Key Insights
The project highlighted several important findings:

• Safety analysis of machine learning perception systems must begin with a clear definition of scope, responsibilities, system interfaces, and hazard traceability 

• Aggregate performance metrics, such as overall accuracy, are insufficient for safety assessment; performance must be evaluated across safety-critical scenarios 

• Dataset bias, annotation drift, and ODD mismatches can introduce latent safety risks that are not visible in standard validation metrics 

• The ML-FMEA provided a shared terminology and framework between machine learning and safety engineering teams. Both teams found this essential to avoid misinterpretation of failure modes and risk severity. 

• Training objectives and evaluation metrics must be explicitly aligned with safety-critical outcomes, as optimization around common driving scenarios can mask weak performance in rare but high-consequence lane configurations 

• Model sensitivity must be balanced across near-field, adjacent-lane, and far-field perception zones, as loss-function weighting can unintentionally create blind spots or over-prioritize distant features 

• Regularization and augmentation techniques should be evaluated with safety intent in mind, as methods that improve robustness to rare lane geometries can also reduce performance in more common nominal driving conditions when applied too aggressively 

Outcome
The application of ML-FMEA improved the consistency and defensibility of the safety analysis by directly linking model development failure modes to the vehicle-level hazards of unintended lane departure and adjacent-lane collision. 

The analysis enabled earlier identification of credible failure modes and mitigations, improved traceability between ML artifacts such as data, labels, and performance metrics, and reduced ambiguity in safety discussions by aligning stakeholders on scope, terminology, and system-level safety impact.