NADS-Net: A Nimble Architecture for Driver and Seat Belt Detection via Convolutional Neural Networks

Sehyun Chun, Nima Hamidi Ghalehjegh, Joseph B. Choi, Chris Schwarz, John G. Gaspar, Daniel V. McGehee, Stephen S. BaekWorkshops @ ICCV 2019 · DOI · arXiv

Abstract

Vehicle safety systems increasingly rely on in-cabin monitoring to detect distracted driving and ensure seat belt compliance. However, existing pose estimation frameworks are often too computationally intensive for real-time deployment on automotive embedded systems. This paper presents NADS-Net, a lightweight, multi-task convolutional neural network (CNN) designed for efficient in-vehicle monitoring. By integrating a Feature Pyramid Network (FPN) backbone with specialized detection heads, NADS-Net simultaneously performs occupant pose estimation and seat belt segmentation. Experimental results on a specialized in-cabin dataset demonstrate that NADS-Net achieves 18 fps on standard hardware with 25% fewer parameters than baseline models, while maintaining high accuracy across diverse lighting conditions.

Introduction

In-Vehicle Monitoring Systems (IVMS) are critical for reducing accidents caused by distracted driving or improper seat belt usage. Despite advancements in computer vision, deploying these systems remains challenging due to two primary factors:

  1. Computational Constraints: High-accuracy models like OpenPose require significant GPU resources, making them unsuitable for the low-power embedded systems typical in vehicles.
  2. Environmental Variability: The cabin environment presents unique challenges, including extreme lighting shifts between day and night, diverse occupant demographics, and the need to detect specific safety equipment like seat belts.

NADS-Net addresses these challenges by providing a "nimble" architecture tailored for the specific geometry and constraints of the automotive interior.

Methodology

The design of NADS-Net prioritizes inference speed without compromising the multi-scale feature extraction necessary for pose estimation.

Architecture

Unlike traditional models that use multi-stage refinement (e.g., VGG-based stacks), NADS-Net employs a Feature Pyramid Network (FPN) backbone based on ResNet-50. This allows the network to extract high-level semantic information while preserving spatial resolution across multiple scales.

NADS-Net Architecture: A ResNet-50 backbone with FPN feeding into three specialized detection heads for key points, part affinity fields, and seat belt segmentation.

The architecture branches into three specialized "heads" from a shared feature map:

  • Key Point Head: Predicts heatmaps for body joints (shoulders, elbows, wrists).
  • Part Affinity Field (PAF) Head: Learns vector fields to associate joints into a coherent human skeleton.
  • Seat Belt Segmentation Head: Generates a binary mask to identify seat belt pixels.

By consolidating these tasks into a single-stage pipeline, the model reduces redundant computations and improves throughput.

Dataset and Evaluation

The model was evaluated using the NADS In-Cabin Dataset, a specialized collection featuring 100 drivers across 50 sessions. The dataset includes 10,550 training and 7,721 testing images, capturing both daytime RGB and nighttime Infrared (IR) footage.

The "Headrest" Metric

Standard pose estimation often uses head size (PCKh) to normalize error. However, in a fixed-camera cabin environment, head sizes vary by individual while the distance to the camera remains relatively constant. The authors proposed using the headrest size as a stable reference point for evaluation, leading to more consistent and reliable performance metrics (mPCKh).

Results and Discussion

NADS-Net demonstrates significant improvements in both efficiency and performance over baseline Part Affinity Field models.

Qualitative results of NADS-Net showing pose estimation and seat belt detection in various lighting and occupant scenarios.

Performance Metrics

  • Efficiency: The model utilizes approximately 39.3 million parameters—a 25% reduction compared to the baseline—while adding seat belt detection capabilities.
  • Speed: Inference reached 18 fps on an NVIDIA 1080Ti, representing a 50% speedup over the 12 fps baseline.
  • Accuracy: NADS-Net achieved an 84% mPCKh, outperforming the baseline's 82%. Seat belt detection showed high specificity (99.28%), though the Intersection over Union (IOU) was lower (46.57%) due to the model predicting thinner masks than human annotators.

Key Insights

  • Contextual Learning: Saliency maps indicated that the model uses facial cues to help localize wrists, suggesting it learns the global context of driving postures.
  • IR Robustness: While initial performance on IR data was lower, fine-tuning on the in-cabin dataset allowed the model to leverage geometric similarities between RGB and IR, proving the effectiveness of transfer learning for nighttime monitoring.

Conclusion and Future Work

NADS-Net provides a robust framework for real-time in-cabin safety monitoring. Future research will focus on addressing demographic biases observed in the dataset (particularly for individuals with darker skin tones or varying hairstyles) and incorporating temporal logic to improve detection stability across video sequences.

Summary updated

BibTeX
@inproceedings{chun2019nads,
  author = {Sehyun Chun and Nima Hamidi Ghalehjegh and Joseph B. Choi and Chris Schwarz and John G. Gaspar and Daniel V. McGehee and Stephen S. Baek},
  title = {{NADS-Net: A Nimble Architecture for Driver and Seat Belt Detection via Convolutional Neural Networks}},
  booktitle = {Workshops @ ICCV},
  year = {2019},
  doi = {10.1109/iccvw.2019.00295},
}
Figure