CNN based lane detection with instance segmentation in edge-cloud computing

Wang, Wei; Lin, Hui; Wang, Junshu · 2020 · DOAJ

DOI: 10.1186/s13677-020-00172-z

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of achieving high-accuracy, real-time lane detection for autonomous driving systems by integrating convolutional neural networks (CNNs) with an edge-cloud computing architecture. Traditional lane detection methods rely on manual feature extraction and are susceptible to environmental variations, while pure cloud-based deep learning approaches suffer from high latency and bandwidth constraints. The authors propose a dual-model approach based on instance segmentation that leverages edge computing for low-latency data processing and cloud computing for heavy computational loads, thereby improving both efficiency and robustness. The methodology employs a two-branch CNN network: a lane segmentation branch that outputs binary maps distinguishing lanes from the background, and a lane embedding branch that clusters pixels into distinct lane instances using a custom loss function involving variance and distance terms. To handle complex road geometries, particularly slope changes, the system utilizes a custom neural network to dynamically generate an inverse perspective transformation matrix. This replaces fixed transformation matrices, allowing the system to project images into a "bird’s-eye view" where lane pixels are fitted using third-order polynomials via least squares. The framework incorporates edge computing to process data closer to the source, reducing transmission delays and enhancing privacy, while utilizing cloud resources for intensive training and processing. Experiments were conducted using the TuSimple dataset, which includes diverse scenarios such as varying lighting conditions and multi-lane configurations. The model was trained with specific hyperparameters, including a batch size of 6 and an exponential decay learning rate strategy. The results demonstrate that the proposed method achieves superior recognition efficiency and accuracy compared to other lane recognition models. Specifically, the dynamic inverse perspective transformation effectively resolves lane line deviation issues in gradient road scenes, and the instance segmentation approach successfully handles dynamic lane changes and multi-lane scenarios. The integration of edge-cloud computing significantly reduces latency and improves the real-time performance of the detection system. The significance of this work lies in its effective combination of deep learning instance segmentation with distributed computing paradigms. By addressing the limitations of fixed geometric models and the latency of centralized cloud processing, the proposed framework offers a robust solution for autonomous driving applications. It enhances the system's ability to adapt to varying road conditions and lighting, providing a scalable and efficient architecture for real-time lane detection that balances computational load and processing speed.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms