Enhancing Cross-Dataset Performance of Distracted Driving Detection With Score Softmax Classifier And Dynamic Gaussian Smoothing Supervision

TL;DR

This work tackles cross-dataset generalization in distracted driving detection under background noise by introducing Score-Softmax (S-Softmax) and Dynamic Gaussian Smoothing Supervision (DGSS), along with a Gaussian fusion strategy for multi-channel score aggregation. The approach untangles the one-hot constraint, injects dynamic label smoothing, and fuses scores across channels via Gaussian statistics, yielding substantial cross-dataset gains without altering backbone architectures. Across SFDDD, AUCDD, 100-Driver, EZZ2021, and HNUDDC1, the method improves robustness to distribution shifts and noise, with notable improvements over several state-of-the-art baselines. The combination of S-Softmax, DGSS, and GF offers a practical path toward more reliable driver monitoring in naturalistic driving conditions, though viewpoint variations remain a challenging gap for future work potentially addressable by integrating CLIP or multimodal cues.

Abstract

Deep neural networks enable real-time monitoring of in-vehicle drivers, facilitating the timely prediction of distractions, fatigue, and potential hazards. This technology is now integral to intelligent transportation systems. Recent research has exposed unreliable cross-dataset driver behavior recognition due to a limited number of data samples and background noise. In this paper, we propose a Score-Softmax classifier, which reduces the model overconfidence by enhancing category independence. Imitating the human scoring process, we designed a two-dimensional dynamic supervisory matrix consisting of one-dimensional Gaussian-smoothed labels. The dynamic loss descent direction and Gaussian smoothing increase the uncertainty of training to prevent the model from falling into noise traps. Furthermore, we introduce a simple and convenient multi-channel information fusion method;it addresses the fusion issue among arbitrary Score-Softmax classification heads. We conducted cross-dataset experiments using the SFDDD, AUCDD, and the 100-Driver datasets, demonstrating that Score-Softmax improves cross-dataset performance without modifying the model architecture. The experiments indicate that the Score-Softmax classifier reduces the interference of background noise, enhancing the robustness of the model. It increases the cross-dataset accuracy by 21.34%, 11.89%, and 18.77% on the three datasets, respectively. The code is publicly available at https://github.com/congduan-HNU/SSoftmax.

Figures (12)

• Figure 1: Similar noise features in in-vehicle camera samples can lead to local traps in the solution space. Prominent sources of noise, such as the windows, rearview mirrors, and the vehicle control panel, show significant variations across datasets. These differences are the primary causes for the local noise traps shown in Fig. \ref{['Figure_optimize']}.
• Figure 2: The first row depicts the details of the weighted summation of scores. Blue circles represent individual neuron outputs, each corresponding to different score weights. Each group corresponds to one category. When weighted and summed according to the score table, this output $\mathcal{S}$ yields the score of $\boldsymbol{X}$ for all categories it could belong to. The second row shows three different designs of two-dimensional supervision matrices, from left to right: $\mathcal{Y}_1$, $\mathcal{Y}_2$, and $\mathcal{Y}_3$, where the supervision strength gradually weakens.
• Figure 3: Overview of the training process. The training comprises three main stages: Pretrain Process, Gaussian Generation, and Finetuning Process. In Pretrain Process, we initialize the network with ImageNet dataset training using Softmax classifier and cross-entropy loss $\mathcal{L}_{ce}$. Gaussian Generation involves creating a Gaussian distribution-guided score supervision matrix, a key contribution of our work. In Finetuning Process, we fine-tune the model on the combined distracted driving dataset, transferring pretrained backbone network weights, and employing the S-Softmax classifier with a score loss function $\mathcal{L}_s$. Class scores are computed using weighted summation (represented by $\Sigma$), and weighted product is represented by $\otimes$.
• Figure 4: Multi-channel information fusion method based on Gaussian fusion. $\mathcal{S}^1$, $\mathcal{S}^2$ mean the predict score matrix of channl D1 and D1, respectively. $\mathcal{S}^F$ is the fusion score matrix regenerated based on Gaussian distribution.
• Figure 5: The first and second rows illustrate attention heatmaps of ResNet18 using Vanilla Softmax with One-Hot labeling, and S-Softmax with DGSS, respectively. These heatmaps are generated through Grad-CAM selvaraju_grad-cam_2020. Noise features can create local traps within the solution space. Softmax might lead CNNs to be overly confident, taking shortcuts that could potentially result in falling into traps. To enhance understanding, the brief schematic of the optimization process in the third row vividly demonstrates how DGSS facilitates loss vibration decline, thereby avoiding noise traps.