LNet: Lightweight Network for Driver Attention Estimation via Scene and Gaze Consistency

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Automotive Engineering

Background:

• Establishing consistency between multi-view scenes and driver gaze is difficult in resource-constrained vehicle systems.
• Existing methods often rely on unidirectional data fusion, leading to high computational loads due to high-resolution image processing and complex semantic extraction.

Purpose of the Study:

• To propose a lightweight driver-attention estimation framework.
• To leverage geometric consistency between scene and gaze for bidirectional feature extraction.
• To improve the accuracy-efficiency trade-off in driver attention monitoring.

Main Methods:

• A lightweight feature extraction module with dual asymmetric branches for parallel global and local information capture.
• An information cross-fusion module to enhance interaction between scene and gaze streams.
• A multi-branch architecture for multi-scale extraction of gaze and geometric cues, reducing computational redundancy.

Main Results:

• The proposed framework achieves a better accuracy-efficiency trade-off compared to prior methods.
• Incorporating scene information does not introduce significant computational overhead.
• Experimental results on a large public dataset validate the framework's effectiveness.

Conclusions:

• The lightweight framework effectively estimates driver attention by leveraging geometric consistency between scene and gaze.
• The approach offers a computationally efficient solution for driver monitoring systems.
• Preliminary exploration suggests potential for predicting attention trends using temporal gaze continuity.