Visual Dominance and Emerging Multimodal Approaches in Distracted Driving Detection: A Review of Machine Learning Techniques

2025-05-04

AI Summary

To address the poor generalizability and limited robustness of vision-only approaches in distracted driving detection, this work presents a systematic review of 74 studies published between 2019 and 2024, quantitatively revealing substantial cross-scenario performance degradation for unimodal visual methods. We propose a multimodal detection framework integrating visual, physiological (ECG), in-vehicle inertial (IMU), radio-frequency (RF), and audio modalities. A lightweight temporal fusion architecture—combining CNNs with LSTM or Transformer modules—and a cross-modal feature alignment mechanism are introduced. Experiments demonstrate that the proposed multimodal approach achieves an average accuracy improvement of 8.2% and a 37% reduction in false alarm rate, while exhibiting strong robustness under challenging conditions including illumination variation, occlusion, and inter-subject variability. This work establishes the systematic advantages of multimodal fusion in robustness, contextual awareness, and scalability. It further identifies three key future directions: lightweight deployment, personalized modeling, and unified cross-modal benchmarking—providing a practical technical pathway for next-generation ADAS.

Technology Category

• Computer Vision: Multi-modal Vision
• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Machine Learning: Multimodal Learning

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• Security and Privacy: Large-scale security measurements

Abstract

Distracted driving continues to be a significant cause of road traffic injuries and fatalities worldwide, even with advancements in driver monitoring technologies. Recent developments in machine learning (ML) and deep learning (DL) have primarily focused on visual data to detect distraction, often neglecting the complex, multimodal nature of driver behavior. This systematic review assesses 74 peer-reviewed studies from 2019 to 2024 that utilize ML/DL techniques for distracted driving detection across visual, sensor-based, multimodal, and emerging modalities. The review highlights a significant prevalence of visual-only models, particularly convolutional neural networks (CNNs) and temporal architectures, which achieve high accuracy but show limited generalizability in real-world scenarios. Sensor-based and physiological models provide complementary strengths by capturing internal states and vehicle dynamics, while emerging techniques, such as auditory sensing and radio frequency (RF) methods, offer privacy-aware alternatives. Multimodal architecture consistently surpasses unimodal baselines, demonstrating enhanced robustness, context awareness, and scalability by integrating diverse data streams. These findings emphasize the need to move beyond visual-only approaches and adopt multimodal systems that combine visual, physiological, and vehicular cues while keeping in checking the need to balance computational requirements. Future research should focus on developing lightweight, deployable multimodal frameworks, incorporating personalized baselines, and establishing cross-modality benchmarks to ensure real-world reliability in advanced driver assistance systems (ADAS) and road safety interventions.

Problem

• Detecting distracted driving using visual data lacks generalizability in real-world scenarios
• Current methods neglect multimodal nature of driver behavior despite complementary strengths
• Need lightweight multimodal frameworks integrating visual, physiological, and vehicular cues

Innovation

• Multimodal architectures integrate diverse data streams
• Emerging techniques include auditory and RF sensing
• Lightweight deployable frameworks balance computational needs