[Paper Review] Applying Semantic Segmentation to Autonomous Cars in the Snowy Environment

Zhaoyu Pan, Takanori Emaru+2 more|arXiv (Cornell University)|Jul 25, 2020

Advanced Neural Network Applications6 references8 citations

TL;DR

This paper proposes a semantic segmentation approach using Fully Convolutional Networks (FCN) to enhance environment perception in snowy conditions for autonomous vehicles. Trained on a custom dataset, the method achieves improved object classification in snow-covered scenes, though results indicate the need for dataset optimization and better algorithms for improved performance.

ABSTRACT

This paper mainly focuses on environment perception in snowy situations which forms the backbone of the autonomous driving technology. For the purpose, semantic segmentation is employed to classify the objects while the vehicle is driven autonomously. We train the Fully Convolutional Networks (FCN) on our own dataset and present the experimental results. Finally, the outcomes are analyzed to give a conclusion. It can be concluded that the database still needs to be optimized and a favorable algorithm should be proposed to get better results.

Motivation & Objective

• To improve environment perception in snowy conditions for autonomous vehicles.
• To address the challenge of semantic segmentation in low-visibility, snow-covered environments.
• To evaluate the performance of Fully Convolutional Networks (FCN) on a custom dataset collected under snowy conditions.
• To identify limitations in current datasets and models for snowy autonomous driving scenarios.

Proposed method

• Training a Fully Convolutional Network (FCN) on a custom dataset collected in snowy environments.
• Using end-to-end deep learning to classify pixels into semantic categories such as road, vehicle, pedestrian, and snow.
• Collecting real-world driving data in snowy conditions to build a domain-specific dataset.
• Employing standard semantic segmentation pipelines with data augmentation to improve robustness.
• Evaluating model performance using standard segmentation metrics on the custom dataset.
• Analyzing failure cases and visualizing segmentation outputs to identify dataset and model limitations.

Experimental results

Research questions

• RQ1 How effective is semantic segmentation using FCN in autonomous driving under snowy conditions?
• RQ2 What are the main challenges in segmenting objects in snow-covered scenes using deep learning?
• RQ3 How does performance on a custom snowy dataset compare to general-purpose datasets?
• RQ4 What limitations exist in current datasets and models for snowy autonomous driving?
• RQ5 What improvements are needed in data collection and model architecture for better generalization in snowy environments?

Key findings

• The proposed FCN-based semantic segmentation model achieves measurable performance in snowy driving conditions, demonstrating feasibility.
• The model struggles with fine-grained object boundaries and class confusion, especially between snow and similar-colored objects.
• Results indicate that the current dataset lacks sufficient diversity and quality for optimal model generalization.
• Visual analysis reveals frequent misclassification of snow-covered road surfaces and obscured vehicles.
• The study concludes that further dataset optimization and algorithmic improvements are necessary for robust performance.
• The authors recommend developing specialized architectures and data augmentation techniques tailored for snowy environments.