ASC-SW: Atrous strip convolution network with sliding windows for visual-assisted map navigation

2025-07-16

AI Summary

Traditional LiDAR sensors struggle to detect ground-level linear obstacles—such as low-hanging power lines—posing significant safety risks to map-based mobile robot navigation. To address this, we propose ASC-SW, a lightweight vision-assisted segmentation framework built upon MobileNetV2. Its core innovation is the Atrous Strip Convolution Spatial Pyramid Pooling (ASCP) module, which enhances feature representation for deformable, slender structures. Additionally, we introduce a sliding-window post-processing mechanism that substantially improves segmentation robustness and accuracy with negligible computational overhead. Evaluated on a custom-built dataset, ASC-SW achieves 75.3% mean Intersection-over-Union (mIoU). Deployed on a Jetson Orin Nano edge device, it operates at 9.3 FPS in real time. Extensive experiments on a physical robot platform confirm its effectiveness in improving navigation safety by reliably detecting hazardous linear obstacles.

Technology Category

• Intelligent Robots: Localization, Mapping, and Navigation
• Computer Vision: Segmentation
• Natural Language Processing: Safety and Robustness

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

With the rapid development of lightweight visual neural network architectures, traditional high-performance vision models have undergone significant compression, greatly improving their computational efficiency and energy consumption ratio. This makes them feasible for deployment on resource-constrained edge computing devices. We propose a visual-assisted navigation framework called Atrous Strip Convolution-Sliding Window (ASC-SW), which leverages a depth camera and a lightweight visual neural network to assist map-based mobile robot navigation. This framework compensates for the inability of traditional light detection and range (LiDAR) sensors to detect ground-level obstacles such as ground-level wires. We introduce a lightweight and efficient segmentation model, Atrous Strip Convolution Network (ASCnet), for detecting deformable linear objects (DLOs). MobileNetV2 is used as the backbone network, and Atrous Strip Convolution Spatial Pyramid Pooling (ASCSPP) is designed to extract DLO features more effectively. Atrous Strip Convolution is integrated into ASCSPP to accurately identify the linear structure of DLOs with low computational cost. Additionally, a Sliding Window (SW) post-processing module is proposed to denoise the output in complex environments, improving recognition accuracy. Our method strikes a balance between inference speed and segmentation performance. It achieves a mean Intersection over Union (Miou) score of 75.3% on a self-built dataset and reaches 9.3 FPS inference speed on the Jetson Orin Nano edge device. Overall, our approach outperforms existing DLO detection models and has been successfully validated on a physical robotic platform.

Problem

• Detects ground-level obstacles like wires for robot navigation
• Improves segmentation of deformable linear objects efficiently
• Balances speed and accuracy for edge device deployment

Innovation

• Lightweight ASCnet for DLO segmentation
• Atrous Strip Convolution for linear feature extraction
• Sliding Window module for denoising output