ML-PersRef: A Machine Learning-based Personalized Multimodal Fusion Approach for Referencing Outside Objects From a Moving Vehicle

TL;DR

This paper introduces a machine learning-based multimodal fusion system for referencing outside objects from a moving vehicle, demonstrating improved accuracy and personalization capabilities in simulated driving scenarios.

Contribution

It presents a novel multimodal fusion approach for vehicle-object referencing and a transfer-of-learning personalization technique for individual driver adaptation.

Findings

Multimodal fusion outperforms single-modality methods.

Personalization improves referencing accuracy for individual drivers.

The approach is effective in simulated driving environments.

Abstract

Over the past decades, the addition of hundreds of sensors to modern vehicles has led to an exponential increase in their capabilities. This allows for novel approaches to interaction with the vehicle that go beyond traditional touch-based and voice command approaches, such as emotion recognition, head rotation, eye gaze, and pointing gestures. Although gaze and pointing gestures have been used before for referencing objects inside and outside vehicles, the multimodal interaction and fusion of these gestures have so far not been extensively studied. We propose a novel learning-based multimodal fusion approach for referencing outside-the-vehicle objects while maintaining a long driving route in a simulated environment. The proposed multimodal approaches outperform single-modality approaches in multiple aspects and conditions. Moreover, we also demonstrate possible ways to exploit behavioral differences between users when completing the referencing task to realize an adaptable personalized system for each driver. We propose a personalization technique based on the transfer-of-learning concept for exceedingly small data sizes to enhance prediction and adapt to individualistic referencing behavior. Our code is publicly available at https://github.com/amr-gomaa/ML-PersRef.