ASC-SW: Visual Navigation Framework for DLO Detection

Updated 6 July 2026

The paper introduces ASC-SW, a navigation framework that integrates a lightweight segmentation model (ASCnet) with a temporal sliding-window module to detect deformable linear objects.
It enhances traditional mapping by converting RGB-segmented DLO regions into depth-filtered obstacle point clouds for costmap integration.
The framework, validated on edge devices, achieves up to 75.3% mIoU and robust real-time performance in cluttered, complex environments.

Searching arXiv for the specified ASC-SW paper and closely related work on atrous convolution, strip convolution, and sliding-window formulations. Atrous Strip Convolution-Sliding Window (ASC-SW) denotes a visual-assisted navigation framework for map-based mobile robots that combines a lightweight semantic segmentation network, Atrous Strip Convolution Network (ASCnet), with a temporal Sliding Window (SW) post-processing module to detect deformable linear objects (DLOs) such as ground-level wires that are often missed by standard 2D LiDAR. The framework uses a depth camera, converts segmented DLO regions into obstacle point clouds for costmap integration, and is reported to achieve a mean Intersection over Union score of 75.3% on a self-built dataset and 9.3 FPS on the Jetson Orin Nano edge device (Liu et al., 17 Jul 2025).

1. Definition and problem setting

In the usage established by "ASC-SW: Atrous strip convolution network with sliding windows for visual-assisted map navigation" (Liu et al., 17 Jul 2025), ASC-SW is not merely a convolutional operator. It is a complete navigation framework with two main algorithmic parts: ASCnet, a lightweight semantic segmentation model for DLO detection, and SW, a temporal post-processing module that denoises the segmentation output in complex environments.

The target failure mode is specific. Standard 2D LiDAR usually scans in a horizontal plane at a fixed height, so ground-level wires or very thin DLOs close to the floor can fall below the LiDAR plane or lack sufficient thickness and reflectivity for robust detection. ASC-SW addresses this by combining RGB segmentation, depth filtering, and costmap injection, thereby augmenting a conventional map-based mobile robot navigation stack with obstacle information that LiDAR typically misses.

A concise decomposition of the framework is given below.

Component	Function	Main elements
ASCnet	DLO segmentation	MobileNetV2, ASCSPP, CA
SW	Temporal denoising	Erosion, contour IDs, window voting
Navigation integration	Obstacle insertion	Masked depth, voxel filter, costmap

This framing is important because the term "sliding window" in ASC-SW refers to temporal denoising over mask sequences, not only to convolutional evaluation. Likewise, "Atrous Strip Convolution" refers to a DLO-oriented feature extractor embedded inside ASCnet rather than to a standalone generic layer.

The system pipeline proceeds from sensing to obstacle-aware planning. A depth camera, exemplified in the implementation by the Intel RealSense D455, provides RGB images and registered depth. ASCnet processes the RGB stream and outputs a binary DLO-versus-background mask. The SW module then operates on the mask stream over time and produces a cleaner, temporally stabilized mask (Liu et al., 17 Jul 2025).

That mask is subsequently used to filter the depth point cloud. Only points whose pixels are labeled as DLO are retained, yielding a DLO point cloud. A voxel filter, specifically a PCL voxel grid, down-samples this point cloud to reduce computation. The filtered DLO point cloud is then injected into the robot’s costmap as obstacle data. Classical map-based navigation components—occupancy grid or costmap, local planner, global planner, and associated control logic—can therefore treat floor-level wires as occupied cells and avoid them in the same way as more conventional obstacles.

The framework is explicitly additive rather than replacement-oriented. LiDAR continues to detect ordinary obstacles, while ASC-SW supplies additional ground-level obstacle information derived from vision and depth. The result is a hybrid navigation stack in which DLO perception is attached to standard map-based navigation rather than embedded into a bespoke end-to-end planner.

3. ASCnet and Atrous Strip Convolution Spatial Pyramid Pooling

ASCnet is the segmentation model at the core of ASC-SW. It is based on DeepLabv3+ and follows an encoder-decoder structure with MobileNetV2 as the backbone network. MobileNetV2 is used with pre-trained weights and contributes the lightweight profile of the overall model through depthwise separable convolutions and inverted residual blocks with linear bottlenecks. In the reported configuration, training uses $224 \times 224$ inputs, and ASCnet has 6.9M parameters and 6.36 G FLOPs, which is close to DeepLabv3Plus-MobileNetV2 at 6.8M parameters and 6.34 G FLOPs (Liu et al., 17 Jul 2025).

The encoder contains the backbone feature extractor, the Atrous Strip Convolution Spatial Pyramid Pooling (ASCSPP) module, and a Channel Attention (CA) module. The decoder fuses low-level and high-level features through a skip-connection style design and upsamples the result to a full-resolution binary segmentation mask. High-level features are processed by ASCSPP, then re-weighted by CA, and finally fused with low-level features before final prediction.

ASCSPP is the defining modification. Classic ASPP in DeepLabv3+ uses multiple parallel atrous convolutions with different dilation rates. ASCSPP replaces those atrous square convolutions with atrous strip convolutions. The module keeps a parallel architecture, uses strip convolutions of the same size in parallel with different dilation rates, and adds the original low-level features back before output to compensate for lost spatial detail. The paper does not specify the exact number of branches, exact kernel lengths, or explicit dilation values; it specifies the design principle and reports its empirical effects.

The rationale is geometric. DLOs are long, thin, and line-like, so strip kernels are aligned with their morphology in a way that square kernels are not. The paper describes vertical and horizontal strip convolutions as cascaded replacements for traditional square kernels and states that Atrous Strip Convolution is integrated into ASCSPP to identify the linear structure of DLOs with low computational cost.

The CA module is based on Squeeze-and-Excitation-style channel weighting, but replaces fully connected layers and ReLU with $1 \times 1$ convolutions to reduce computation. In the ablation study, CA improves mIoU from 0.6099 to 0.6367 in Scenario 1 and from 0.4291 to 0.5069 in Scenario 2. The ASCSPP ablation is similarly explicit: ASPP yields mIoU $= 0.4386$ at 223 FPS, DAPP yields 0.4906 at 205.92 FPS, SPASPP yields 0.336 at 205.47 FPS, and ASCSPP yields 0.5069 at 217 FPS. These results place ASCSPP as the best-performing spatial-pyramid variant in the reported comparison while preserving nearly the same throughput as the simplest ASPP configuration.

4. Sliding Window post-processing and temporal consistency

The SW module is a temporal denoising algorithm applied after framewise segmentation. Its purpose is to suppress intermittent false positives and preserve the most temporally consistent DLO object in cluttered scenes, particularly when camera viewpoints differ from the top-down images used during training and when line-like distractors such as baseboards or floor patterns are present (Liu et al., 17 Jul 2025).

The algorithm begins with a morphological erosion step:

$M_{\text{erode}} \leftarrow Erode(M,\text{Kernel}(m,n)),$

with kernel size $(m,n) = (1,1)$ in the reported implementation. Erosion is used to repair fragmented mask regions and remove small isolated pixels. Connected components are then extracted from $M_{\text{erode}}$ , and contours with area greater than 50 pixels are assigned IDs.

ID assignment is based on centroid tracking. For contours $[C_1,\ldots,C_n]$ , the algorithm computes

$[Cen_1,\ldots,Cen_n] \leftarrow Centroid([C_1,\ldots,C_n]).$

Current centroids are matched to those from the previous frame by Euclidean distance. If the distance is below a threshold of 50 pixels, the contour inherits a tracked ID; otherwise it receives a new ID.

The sliding-window voting stage maintains a queue of the IDs observed in the last $k = 45$ frames. The most frequent ID in that queue is selected, and only the region corresponding to that ID is retained:

$M' \leftarrow Filter(M_{\text{erode}}, [ID_n]).$

This mechanism favors persistent detections and suppresses frame-to-frame flicker. In the reported experiments, SW improves ASCnet from 0.6367 to 0.7533 mIoU in Scenario 1 and from 0.5069 to 0.7428 in Scenario 2. The same post-processing also improves other segmentation backbones, although the paper notes that SW is not applied to RT-DLO-640×360 in the comparison table because RT-DLO already has few false positives and SW can worsen results when the predicted DLO mask is incomplete.

The module therefore acts as a temporal regularizer rather than as a learned recurrent component. Its operation is deterministic, geometry-aware through centroid tracking, and lightweight enough that on an RTX 3070 it produces no measurable FPS drop relative to framewise segmentation alone.

5. Training protocol, evaluation, and deployment

The training data are drawn from two sources. The first is an auto-generated wires dataset with 7200 RGB images at resolution $1 \times 1$ 0, split into 5760 training images and 1440 validation images. The second is a self-built dataset with 1850 RGB images at the same resolution, split into 1480 training images and 370 validation images, collected to compensate for the lack of white and black DLO examples in the auto-generated set. In the self-built set, the authors used SAM to segment DLOs and replaced backgrounds with 10 different backgrounds for augmentation (Liu et al., 17 Jul 2025).

All training images were captured from a top-down perspective, perpendicular to the plane where the DLO lies. The evaluation set is described as a self-built testing dataset of 548 real-world images at $1 \times 1$ 1, captured with a RealSense D455 and organized into two scenarios with varying floor colors and DLO colors designed to stress cases in which DLO appearance is similar to the background.

Training uses Ubuntu 20.04, Python 3.8, and an NVIDIA RTX 3070 with 8 GB VRAM. The optimizer is Adam with learning rate $1 \times 1$ 2, batch size 4, and 200 epochs. The backbone is initialized with ImageNet pre-trained weights. The paper defines the evaluation metrics as

$1 \times 1$ 3

$1 \times 1$ 4

$1 \times 1$ 5

$1 \times 1$ 6

$1 \times 1$ 7

The reported computational profile is central to the framework’s intended use on edge devices. On an RTX 3070 with $1 \times 1$ 8 input, DeepLabv3Plus-ResNet101 has 59.9M parameters, 14.53 G FLOPs, 10.04 ms latency, and 157 FPS; DeepLabv3Plus-MobileNetV2 has 6.8M parameters, 6.34 G FLOPs, 3.41 ms latency, and 223 FPS; ASCnet has 6.9M parameters, 6.36 G FLOPs, 3.51 ms latency, and 217 FPS. On the Jetson Orin Nano, DeepLabv3Plus-MobileNetV2 reaches 9.48 FPS, ASCnet reaches 9.34 FPS, DeepLabv3Plus-MobileNetV2+SW reaches 9.46 FPS, and ASCnet+SW reaches 8.81 FPS. On the Jetson TX1, ASCnet reaches 7.38 FPS and ASCnet+SW reaches 6.3 FPS.

The segmentation comparisons show a consistent trade-off between model size, input resolution, and temporal post-processing. Before SW, RT-DLO-640×360 yields the best mIoU in both reported scenarios, while ASCnet is second best and substantially stronger than DeepLabv3Plus-MobileNetV2 at the same low input resolution. After SW, DeepLabv3Plus-ResNet101+SW achieves 0.767 mIoU in Scenario 1, ASCnet+SW achieves 0.75336, and ASCnet+SW becomes the best model in Scenario 2 with 0.74277 mIoU. The framework has also been validated on a physical mobile robot equipped with a depth camera, Jetson-class edge computing, and a map-based navigation stack, where the clean DLO masks are converted to point-cloud obstacles and incorporated into costmap-based planning.

The exact acronym ASC-SW is context-specific rather than standardized. Its constituent ideas derive from earlier lines of work on atrous convolution, strip-shaped kernels, and sliding-window formulations. DeepLab defined atrous convolution explicitly as

$1 \times 1$ 9

with effective kernel size

$= 0.4386$ 0

thereby enlarging the field of view without increasing the number of parameters or the amount of computation. DeepLab also introduced atrous spatial pyramid pooling (ASPP), in which multiple parallel atrous branches with different rates capture multi-scale context (Chen et al., 2016). ASCSPP in ASCnet is best understood as a DLO-oriented specialization of that logic, replacing square atrous branches with strip-shaped ones.

A second strand appears in time-series forecasting. "Time Series Forecasting via Semi-Asymmetric Convolutional Architecture with Global Atrous Sliding Window" formulates a global atrous sliding window through a Central-Bidirectional Atrous Algorithm (CBAA) operating over a 2D strip representation built from a differenced series and a fuzzy-inspired universe of discourse. It then applies Semi-Asymmetric Convolution (SAC) with horizontal and vertical strip kernels of shapes $= 0.4386$ 1 and $= 0.4386$ 2 to couple global-value context and local-window structure (He, 2023). In that setting, "sliding window" denotes windowed sequence segmentation, while "strip convolution" denotes anisotropic filtering over a constructed 2D feature map.

A later forecasting paper, "Adaptive Fuzzy Time Series Forecasting via Partially Asymmetric Convolution and Sub-Sliding Window Fusion," extends the same family of ideas. It defines a Bilateral Atrous Algorithm (BAA) that performs atrous convolution separately on the left and right context around a preserved center element, and it applies partially asymmetric convolutions with $= 0.4386$ 3 and $= 0.4386$ 4 sequentially to create sub-windows within reconstructed windows (Li, 28 Jul 2025). This suggests a broader design pattern in which atrous sampling enlarges contextual reach and strip kernels impose directional bias.

A third related strand is light-field compression. "LFIC-DRASC: Deep Light Field Image Compression Using Disentangled Representation and Asymmetrical Strip Convolution" introduces Asymmetrical Strip Convolution (ASC) through a strip convolution module that combines Conv $= 0.4386$ 5, Conv $= 0.4386$ 6, and Conv $= 0.4386$ 7, followed by $= 0.4386$ 8 fusion and a residual connection. For $= 0.4386$ 9, the layer uses Conv $M_{\text{erode}} \leftarrow Erode(M,\text{Kernel}(m,n)),$ 0, Conv $M_{\text{erode}} \leftarrow Erode(M,\text{Kernel}(m,n)),$ 1, Conv $M_{\text{erode}} \leftarrow Erode(M,\text{Kernel}(m,n)),$ 2, and Conv $M_{\text{erode}} \leftarrow Erode(M,\text{Kernel}(m,n)),$ 3 to capture long-range correlation in light-field feature space (Feng et al., 2024). That work does not use atrous convolutions or an explicit sliding-window mechanism, but it establishes strip-shaped kernels as an efficient means of modeling elongated or highly anisotropic structure.

Taken together, these papers indicate that the 2025 robotics ASC-SW framework occupies an intersection of three established ideas: atrous receptive-field expansion, strip-oriented anisotropic convolution, and sliding-window temporal or dense-prediction processing. In the robotics formulation, the interaction is distinctive: atrous strip convolutions act inside a lightweight DLO segmentation network, while the sliding window is deferred to a post hoc temporal filtering stage that enforces mask consistency for navigation.