Fast Human Attention Prediction for Fixation-guided Active Perception in Autonomous Navigation

TL;DR

A Liquid Neural Network-based lightweight model rapidly predicts human gaze for active robot perception.

Problem

Existing human attention prediction models are computationally expensive, making them difficult to apply to real-time active perception in autonomous robots.

Approach

The authors designed GazeLNN, using MobileNetV3 as the feature extractor and Liquid Neural Networks as the recurrent engine. The architecture predicts sequential fixation heatmaps auto-regressively, conditioned on the current visual stimulus and fixation history.

Results & Contribution

It achieved state-of-the-art performance on the MIT Low Resolution dataset with a 0.47 ScanMatch score, reducing computational costs by 99.40% and accelerating inference by up to six times compared to existing recurrent baselines. By integrating GazeLNN into a reinforcement learning-trained camera-robot control policy, the practical utility of this efficient architecture for human-fixation-guided perception was validated through real-world aerial robot deployments.