Kaveti Pavan

Low-cost webcam gaze estimation is increasingly proposed as a front-end sensor for self-monitoring of attention and visual well-being, yet average accuracy hides where and for whom such a sensor is reliable. Using only three public gaze datasets (MPIIFaceGaze, Columbia Gaze, and RT-GENE; no new data collection), we make three contributions. First, taking a naive sample-wise split (shuffling samples, not subjects) that places every one of 84 subjects across train/validation/test as a deliberate reference baseline, we show apparent error rising from 4.47 deg on seen subjects to 11.63 deg on subject-disjoint unseen subjects (each a five-seed mean; 95% CI: +/−0.11 deg and +/−4.64 deg, respectively), with an optimism of +7.15 deg (~62%); every naive seed lies far below every subject-disjoint seed, and because the unseen estimate varies widely across seeds, its magnitude should be read as indicative rather than precise. Second, we audit reliability per subject and head pose stratum without demographic labels, revealing a 4.9× spread across users and a Monte Carlo dropout uncertainty that provides a partial, borderline per-subject reliability signal (Pearson r = 0.48; Spearman rho = 0.50, with a bootstrap interval including zero); bias-only recalibration lowers error only modestly (subject-macro: 11.79 to 11.50 deg) and does not help the least-reliable users. Third, we convert this measured reliability profile into a per-user decision (trust/recalibrate/abstain) using objective-weight multi-criteria decision making (averaged CRITIC and entropy weights with TOPSIS; the largest weight falls on MC-dropout uncertainty at 0.42, and then pose sensitivity at 0.23); the decision layer disagrees with a single-metric threshold for 35% of users (six of 17) while remaining stable under +/−20% weight perturbation (no action changes). This study provides an exploratory, retrospective, public-data-only implementation of a per-user reliability decision scaffold for low-cost gaze-based self-monitoring.

Driver inattention detection is crucial for road safety, as stress can impair cognitive functions and increase accident risk. Recent advances in wearable technology have led to an increase in the use of multimodal physiological signals for driver inattention detection. Integrating attention mechanisms into these systems has shown promise in enhancing inattention detection. However, attention features can be affected by noise in the data, presenting a significant challenge. To address this, we propose a multimodal differential self-attention-based 1-D convolutional neural network (MDSA-1DCNN) to reduce noise in attention features. In this study, we evaluate the effectiveness of MDSA-1DCNN on multimodal 1-D biosignals obtained from textile electrodes, collecting single-lead electrocardiogram (256 Hz) and respiration (128 Hz) data from 15 healthy participants in two driving states: normal and inattention. The raw data were divided into nonoverlapping segments of 10, 15, and 20 s and preprocessed using the Neurokit Toolbox. These processed segments from each modality were then passed through convolutional blocks to extract temporal features. Self-attention was applied to these features, followed by a differential attention layer to reduce noise. The resulting features were fed into dense layers to classify driver inattention state. The proposed MDSA-1DCNN approach achieved a weighted F-score of 73.23$\%$ and an average accuracy of 78.51$\%$ on the validation set using leave-one-subject-out cross-validation. Future work will explore the utilization of data from multiple sensors and investigate sensor fusion techniques.

Driver inattention detection remains a critical challenge in driver's well-being, requiring robust systems that can distinguish stress-induced mental load during naturalistic driving. Current approaches face limitations in multiple wearable-based data fusion and real-time biosignals assessment. This letter proposes a novel cross-squeeze-and-excitation convolution neural network (CNN) framework to process simultaneously acquired multiple wearable from 15 participants in controlled driving scenarios. The multimodal signals are applied to multistage attention mechanisms [electrocardiogram (ECG)$\leftrightarrow$electrodermal activity (EDA), ECG $\rightarrow$ EDA, EDA $\rightarrow$ ECG] with 1D-CNN blocks, optimized for 10-s signal segments. The proposed approach is able to classify drive inattention state. It is observed that ECG $\rightarrow$ EDA attention achieves 76.54% average accuracy using leave-one-subject-out cross validation, outperforming unimodal approaches by 12.4% and bidirectional attention by 4.8%. Feature visualizations confirm enhanced pattern discrimination in inattention conditions. This letter advances driver health monitoring systems through effective wearable integration and adaptive feature weighting, with potential for edge deployment and clinical stress assessment applications

[Picture 1] Textile sensors enable non-invasive health monitoring, crucial for ensuring road safety by conducting mental well-being checks for drivers. Assessing driver stress with multimodal data from textile electrodes requires effectively integrating and interpreting diverse physiological signals and kinematic data. In this study, we evaluated textile electrodes and cross-attention mechanisms for assessing driver stress using multimodal data. Electrocardiography and Respiration data were collected from 15 healthy volunteers wearing smart shirts in two driving scenarios. Signals were sampled at 256Hz and 128Hz, respectively, with vehicle data also recorded. Segmented Physiological & Vehicle data enter separate networks, 1D Convolutional layers for signals and fully connected layers for vehicle data. Cross-attention fuses Physiological data, these features are combined with vehicle data for stress classification using sigmoid. The proposed approach is able to classify driver stress states using multimodal data, achieving an average accuracy of 79% and an average F-score of 75%. The integration of a cross-attention mechanism facilitates the capture of inter-modality information.

Driver distraction, crucial for road safety, can benefit from multimodal physiological signals assessment. However, fusion of heterogeneous data is highly challenging. In this study, we address this challenge by exploring 1D convolution neural network (CNN) with squeeze and excitation networks (SEcNN) on multimodal data. For this, electrocardiogram (256Hz) and respiration (128Hz) are obtained from subjects (N=10) while using textile electrodes and driving in different scenarios namely normal, texting and calling. The obtained multimodal data is preprocessed and SEcNN to identify driver distraction. Experiments are performed using Leave-one-out-subject cross validation. The proposed approach is able to discriminate drivers distraction. It is observed that SEcNN yields average accuracy 57.03% and average F1 score 54.90% for shorter segments. Thus, the proposed approach using wearable shirts could be useful for non-intrusive monitoring in real world driver scenarios.