Personalised neural networks for a driver intention prediction: communication as enabler for automated driving

Reschke, Johannes; Neumann, Cornelius; Berlitz, Stephan · 2020 · Crossref

DOI: 10.1515/aot-2020-0035

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of establishing trust and communication between pedestrians and automated vehicles. As automation levels increase, drivers become passive passengers, eliminating the informal eye contact and cues pedestrians rely on for safety. To prepare the public for this transition, the authors propose using current manual driving modes to teach pedestrians new light-based signals. This requires a reliable driver intention prediction system that can automatically trigger these signals. The core problem is that general prediction algorithms often suffer from high false-positive rates, which could lead to dangerous situations where a vehicle signals a stop while the driver intends to continue. To solve this, the authors developed a three-stage algorithm for predicting driver intention at pedestrian crossings. Stage one utilizes a recurrent neural network with four LSTM layers to predict time series for five output signals (velocity, brake pressure, steering angle, and accelerations) over a 2-second horizon. Stage two employs a random forest classifier to determine the driver’s willingness to stop based on these predictions. Stage three validates this decision against a velocity threshold of 7 km/h. The system was trained on a dataset of 7,114 time series from multiple drivers, totaling approximately 120 hours of data. The general algorithm achieved a true-positive rate (TPR) of 94.0% with a false-positive rate (FPR) of 5.0%. To improve accuracy and reduce false positives, the study implemented personalization using transfer learning. Data from four specific drivers were collected on reference routes in Ingolstadt, Germany. The researchers tested various transfer learning methods, finding that fine-tuning the last two fully connected layers of the neural network yielded the best results, as it adapted to driver-specific output characteristics while preserving learned temporal dependencies. Personalization improved the TPR for most drivers, ranging from 91.5% to 96.6% at a 5.0% FPR. One driver achieved a TPR of 95.2% with an FPR of 1.8%. However, personalization decreased accuracy for one driver with unpredictable behavior. The study determined that approximately 55 minutes of driving data is optimal for personalization, balancing accuracy gains with training effort. The significance of this work lies in its potential to enable safe vehicle-pedestrian communication in automated driving. By accurately predicting driver intent, vehicles can display intuitive light signals during manual driving, allowing pedestrians to learn these cues before widespread automation. The personalized algorithm outperforms existing literature, achieving higher accuracy than previous studies. However, the authors note that the remaining false-positive rate, though low, poses a safety risk. Future work must focus on further reducing false positives and exploring personalization based on separate driver and vehicle characteristics to support scenarios where multiple users share a vehicle.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• telematics crash prediction

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Theoretical Contribution: computational model