ASC-SW: A Lightweight Atrous Strip Convolution Network for DLOs Segmentation on Edge mobile Robots

Technical Summary: ASC-SW for DLO Segmentation on Edge Mobile Robots

Problem Statement
Detecting Deformable Linear Objects (DLOs), such as floor cables, is critical for the safe navigation of mobile robots. However, existing solutions face significant challenges when deployed on edge mobile platforms. Current state-of-the-art methods are predominantly designed for manipulator platforms operating with fixed, top-down camera views. These models often rely on heavy architectures (e.g., SAM, large ResNets) that are computationally prohibitive for edge devices. Furthermore, a "cross-view" generalization gap exists: models trained on perpendicular manipulator views fail to generalize to the oblique, ground-level perspectives typical of mobile robots, leading to poor performance in complex, noisy environments with varying viewpoints and distractors.

Methodology
The authors propose ASC-SW, a lightweight, geometry-aware segmentation framework designed to bridge the gap between manipulator training data and mobile robot inference. The system comprises two primary components:

1. ASCNet (Atrous Strip Convolution Network):

   • Architecture: Based on the DeepLabV3+ architecture with a MobileNetV2 backbone to ensure lightweight deployment.
   • Atrous Strip Convolution (ASConv): To address the elongated, anisotropic nature of DLOs, the authors replace standard isotropic convolutions with parallel strip convolutions (horizontal and vertical) integrated with atrous (dilated) convolutions. This design enhances sensitivity to linear structures while maintaining a large receptive field without increasing spatial resolution or computational cost.
   • ASCSPP (Atrous Strip Convolution Spatial Pyramid Pooling): A modified pooling module that utilizes parallel atrous strip convolutions with varying dilation rates. Unlike serial stacking methods (e.g., SPASPP, DAPP) that increase parameters, ASCSPP aggregates multi-scale anisotropic features efficiently.
   • Channel Attention (CA): A lightweight attention mechanism using 1×1 convolutions to weight feature maps along the channel dimension, replacing heavier fully connected layers.

2. Sliding Window (SW) Post-Processing:

   • To mitigate false positives caused by noise and low-resolution inputs (224×224), a temporal refinement module is applied.
   • Process: The binary mask output is first eroded to repair fragmentation. Connected components are tracked across frames using centroid matching and Euclidean distance. A sliding window counts the occurrence frequency of each ID; regions with high temporal consistency are retained, while sporadic noise (false positives) is suppressed.

Key Contributions

• Cross-View Formulation: The paper formulates a novel task where models trained exclusively on manipulator-view (top-down) data are adapted to generalize to mobile robot (oblique) viewpoints without requiring additional mobile robot training data.
• Geometry-Aware Module: The introduction of ASCSPP, which integrates atrous and strip convolutions to capture directional sensitivity specific to linear objects while preserving model efficiency.
• Temporal Refinement: A lightweight Sliding Window module that significantly improves robustness against noise and viewpoint-induced errors in dynamic mobile environments.
• Edge Deployment: Validation of the system on physical edge hardware (Jetson TX1 and Orin Nano), demonstrating feasibility for real-world navigation.

Experimental Results
The method was evaluated on a self-built real-world dataset containing 801 images captured by a mobile robot with an Intel RealSense D455 camera, featuring various DLOs (cables, ropes) on different floor backgrounds.

• Performance: ASC-SW achieved a 74.1% mIoU on the mobile robot test set, significantly outperforming baseline models like RT-DLO (18.17% mIoU under the same cross-view conditions) and standard DeepLabV3+ variants.
• Efficiency: The model operates at 261 FPS on an NVIDIA RTX 3070 and remains deployable on edge devices (e.g., 18.45 FPS on Jetson Orin Nano).
• Comparison:
  • ASC-SW outperformed heavy models (e.g., Grounded-SAM, DLOPerceiver) which either failed to run on edge hardware or suffered severe performance degradation when input resolutions were reduced.
  • The Sliding Window module improved the mIoU of ASCNet from 62.2% to 74.05% with only a minor increase in latency (approx. 0.3ms).
  • In cross-view testing, ASC-SW maintained structural continuity where other models produced significant false positives.

Significance and Claims
The paper claims that ASC-SW successfully addresses the limitations of current DLO segmentation by enabling the use of existing manipulator datasets for mobile robot navigation. By aligning the receptive fields with the geometry of DLOs and enforcing temporal stability, the method achieves a balance between lightness and accuracy. The authors emphasize that their approach allows mobile robots to effectively avoid ground obstacles (DLOs) using edge-computing resources.

However, the authors maintain a modest tone regarding future capabilities. They acknowledge that while the system is deployable, the current inference speed on edge devices (approx. 6.45 FPS on Jetson TX1) has not yet reached the standard of real-time performance (30 FPS) required for all high-speed navigation scenarios. Future work is identified as necessary to explore even more efficient segmentation methods.