Personalized driving assistance algorithms: Case study of federated learning based forward collision warning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Automotive Engineering

Background:

• Current advanced driving assistance systems (ADAS) use uniform collision warning algorithms, failing to account for diverse driving behaviors and reducing driver trust.
• Existing personalization methods involve manual threshold adjustments or require impractical large-scale individual data collection.

Purpose of the Study:

• To develop self-adaptive algorithms for personalized forward collision warning (FCW) using federated learning.
• To address the limitations of current ADAS by enhancing driver trust through personalized algorithms.

Main Methods:

• A baseline Long Short-Term Memory (LSTM) model was developed for FCW.
• A federated learning framework was implemented to aggregate knowledge from multiple drivers while preserving privacy.
• A driver-specific batch normalization (BN) layer was integrated into individual vehicle models to manage driving behavior heterogeneity.

Main Results:

• The proposed federated-based personalized models with the BN layer demonstrated superior performance.
• The average modeling accuracy reached 84.88%, comparable to traditional methods using total data collection.
• The addition of the BN layer improved accuracy by an average of 3.48%.

Conclusions:

• Federated learning with a driver-specific BN layer offers an effective and privacy-preserving approach for personalized FCW systems.
• This method enhances driver trust by adapting to individual driving behaviors without compromising data privacy or requiring extensive data collection.