PaperlayerView exact revision on arXiv

Source-linked AI summary

GANav: Efficient Terrain Segmentation for Robot Navigation in Unstructured Outdoor Environments

Tianrui Guan, Divya Kothandaraman, Rohan Chandra, Adarsh Jagan Sathyamoorthy, Kasun Weerakoon, Dinesh Manocha

arXiv:2103.04233v5cs.ROcs.CV

TL;DR

Off-road robots need to identify safe terrain despite ambiguous visual features, while perception must remain accurate and computationally efficient. GANav uses coarse-grained terrain segmentation with group-wise attention and evaluates it on off-road datasets and real robot navigation. It reports improved mIoU across RUGD and RELLIS-3D, along with navigation benefits on Clearpath robots.

  • Problem

    Off-road navigation needs visual classification of terrain by navigability because similar appearances, unclear boundaries, and dangerous-region errors challenge safe perception.

  • Method

    GANav performs coarse-grained semantic segmentation with a transformer-based multi-scale architecture and group-wise attention focused on terrain groups.

  • Results

    1.66−44.35% mIoU improvement is reported on RUGD, while 5.17−19.06% improvement is reported on RELLIS-3D.

  • Takeaways & Limitations

    GANav demonstrates improved accuracy and efficiency over SOTA methods and supports higher-success-rate navigation in real-world unstructured terrains.

  • Takeaways & Limitations

    The evaluation focuses mostly on perception metrics, while label consistency remains challenging because objects such as trees may receive different annotation interpretations.

Abstract

from arXiv · show

We propose GANav, a novel group-wise attention mechanism to identify safe and navigable regions in off-road terrains and unstructured environments from RGB images. Our approach classifies terrains based on their navigability levels using coarse-grained semantic segmentation. Our novel group-wise attention loss enables any backbone network to explicitly focus on the different groups' features with low spatial resolution. Our design leads to efficient inference while maintaining a high level of accuracy compared to existing SOTA methods. Our extensive evaluations on the RUGD and RELLIS-3D datasets shows that GANav achieves an improvement over the SOTA mIoU by 2.25-39.05% on RUGD and 5.17-19.06% on RELLIS-3D. We interface GANav with a deep reinforcement learning-based navigation algorithm and highlight its benefits in terms of navigation in real-world unstructured terrains. We integrate our GANav-based navigation algorithm with ClearPath Jackal and Husky robots, and observe an increase of 10% in terms of success rate, 2-47% in terms of selecting the surface with the best navigability and a decrease of 4.6-13.9% in trajectory roughness. Further, GANav reduces the false positive rate of forbidden regions by 37.79%. Code, videos, and a full technical report are available at https://gamma.umd.edu/offroad/.

I. INTRODUCTION

GANav targets safe terrain identification for autonomous navigation in unstructured outdoor environments, where visual ambiguity and dangerous misclassification complicate perception. It combines coarse-grained segmentation with group-wise attention to improve accuracy and efficiency, then evaluates benefits for real-world navigation.

  • Motivation: Off-road navigation requires visually distinguishing terrain types according to their navigability so robots can identify safe regions.Concrete or asphalt may be highly navigable, whereas rocks or gravel may be bumpy and unsuitable.
  • Results: 10% higher navigation success rate is reported in real-world demonstrations using Clearpath Jackal and Husky robots.The figure passage summarizes this navigation improvement alongside the segmentation results.
  • Motivation: Similar appearances, overlapping boundaries, and small terrain regions can produce wrong classifications, with dangerous or forbidden-region errors carrying serious navigation consequences.These challenges include visually similar classes such as water and puddles.
  • Approach: Coarse-grained segmentation is more appropriate for navigation than fine-grained object labeling when navigation characteristics matter more than individual object identities.For collision avoidance, recognizing trees and poles as obstacles can be sufficient without separately segmenting object types.
  • Approach: GANav combines a multi-scale feature extractor with a transformer-based architecture and group-wise segmentation head to focus on different terrain types.The design is intended to balance segmentation accuracy with computational overhead and supports different feature extractors.
  • Results: 2.25−39.05% mIoU improvement is reported on RUGD, while 5.17−19.06% improvement is reported on RELLIS-3D over prior methods.The contribution passage reports these gains for semantic segmentation and navigable-region classification.

B. Navigation in Outdoor Scenes

Outdoor terrain navigation must infer traversability from uneven, visually ambiguous scenes while managing sensor and computational constraints. GANav addresses this with coarse-grained, efficient attention-based segmentation for navigation-oriented terrain classification.

  • Navigation in Outdoor Scenes: Terrain-navigation methods include proprioceptive, geometric, and appearance-based approaches for estimating terrain features and traversability.Proprioceptive methods use robot vibrations, whereas geometric methods use sensors such as lidar and stereo cameras.
  • Navigation in Outdoor Scenes: Proprioceptive methods require the robot to traverse regions for data collection and assume the robot can navigate across the entire terrain.These assumptions limit their suitability before a route is attempted.
  • Challenges: Recognition and segmentation in unstructured off-road scenes remain challenging because classes such as puddles and asphalt may lack clear boundaries.The supplied passage contrasts this setting with high-accuracy semantic segmentation on standard object and driving datasets.
  • GANav: GANav uses coarse-grained segmentation to classify terrain navigability while addressing similar visual features and unclear object boundaries.The method is designed for off-road environments where fine-grained segmentation methods can degrade.
  • GANav: Self-attention is applied during multi-scale fusion, and generated attention maps contribute to training loss while keeping computational overhead low.The attention-map processing branch is skipped during inference for efficiency.

A. Problem Definition

The problem formulation maps an RGB terrain image and semantic labels into coarse terrain groups and pixelwise group-probability maps. GANav extracts and fuses multi-scale features with attention, explicitly supervises group-specific maps, and reduces inference cost by omitting that branch at test time.

  • A. Problem Definition: The input is an RGB image with pixelwise semantic labels, which are converted into G coarse-grained terrain groups and binary masks.Each binary mask corresponds to one terrain group.
  • A. Problem Definition: The network outputs G-channel probability maps, where each pixel receives a probability distribution over the terrain groups.This defines the coarse-grained semantic segmentation objective.
  • B. Backbone Design: GANav uses a transformer-based backbone to extract and fuse multi-scale features and can adapt the design to different multi-scale backbones.The paper specifically uses a modified Mixed Transformer and also evaluates ResNet50 and Bottleneck Transformer alternatives.
  • C. GANav Segmentation Head: Features from different scales are spatially aligned, resized to a common resolution, concatenated, and passed to multi-head self-attention for fusion.The fused representation is later reshaped and processed by the segmentation head.
  • C. GANav Segmentation Head: Choosing a smaller fusion resolution lowers complexity but can degrade performance, motivating group-wise attention loss to mitigate that drop.The loss is intended to improve overall performance while preserving the efficiency of low-resolution fusion.
  • C. GANav Segmentation Head: Group-wise attention produces one attention map per terrain group and trains each map toward its corresponding binary ground truth using binary cross-entropy.The attention branch is skipped during inference to improve efficiency.

D. Group-wise Attention Loss

Group-wise attention loss guides each attention head toward terrain-group regions while standard segmentation and auxiliary losses optimize the network. Its purpose is intermediate feature guidance, not direct group-distribution prediction.

  • Spatial-resolution reduction can degrade performance, motivating group-wise attention loss to mitigate this effect.
  • Each attention map’s diagonal is reshaped into a spatial map whose pixels represent self-attention scores for one attention head.The map is resized with bi-linear image resizing.
  • Binary cross-entropy guides each attention map using the corresponding group’s binary ground truth.
  • The loss guides which regions an attention head should focus on for later classification rather than accurately predicting group distributions.
  • Multi-class cross-entropy and an auxiliary loss complete training, with parallel deep supervision providing regularization and reducing training time.

E. GA-Nav v.s. Prior Methods

GA-Nav differs from prior transformer-based segmentation methods in computational design and in assigning trainable parameters to terrain groups or semantic classes.

  • SETR and DPT use designs with different computational overhead from GA-Nav, while GA-Nav removes positional embedding from its backbone.
  • Segformer uses no class-specific parameters before final classification, whereas Segmenter and GA-Nav dedicate trainable parameters to classes or groups.
  • Segmenter uses G class embeddings for semantic classes, while GA-Nav uses trainable parameters dedicated to specific groups.

F. GA-Nav-based Segmentation for Robot Navigation

GA-Nav segmentation is converted into a cost map and combined with TERP’s elevation cost map to plan least-cost robot trajectories. Experiments use off-road benchmarks and coarse navigability classes with standard segmentation metrics.

  • GA-Nav’s segmented output is projected onto the ground to create a segmentation cost map before integration with TERP.
  • The segmentation and elevation cost maps are added element-wise, and TERP uses the resulting map to compute least-cost trajectories toward the goal.
  • Experiments use RUGD and RELLIS-3D as off-road terrain benchmarks, with fixed 240K-iteration comparisons and dataset-specific training configurations.RUGD uses batch size 8 and 300 × 375 crops; RELLIS-3D uses batch size 2 and 375 × 600 crops.
  • Evaluation reports IoU, mIoU, mAcc, and aAcc for six coarse-grained classes: smooth, rough, bumpy, forbidden, obstacle, and background.

B. Ablation Studies

Ablations show benefits from multi-scale fusion and group-wise attention across backbones, while qualitative results assess navigable-region detection and robustness to similarly colored objects.

  • Multi-scale Fusion Design: 14.76% and 11.43% mIoU improvements result from the multi-scale fusion design on RUGD and RELLIS-3D, respectively.
  • Qualitative Results: Qualitative comparisons organize each row as the original image, ground truth, two SOTA predictions, and GA-Nav prediction.
  • Qualitative Results: GA-Nav detects different navigable regions and is robust to objects with similar colors in the qualitative comparisons.
  • Group-wise Attention Design: 6.67−10.85% and 4.18−10.31% mIoU improvements from group-wise attention occur across backbone choices on RUGD and RELLIS-3D, respectively.
  • Ablation on Spatial Reduction with GA Head: Without group-wise attention heads, changing resolution to 1/16 and 1/32 causes performance drops of 1.3% and 4.0%.

C. Run-time Analysis and Computational Cost

GANav combines competitive segmentation accuracy with low computational cost. Its group-wise attention improves performance across backbones and mitigates accuracy loss from spatial reduction while retaining efficient inference.

  • Inference Efficiency: GANav maintains state-of-the-art mIoU even when GANav-r16 and GA-Nav-r32 trade performance for faster inference.The comparison highlights efficiency without abandoning competitive accuracy.
  • Backbone Ablation: 7.26−11.44% and 4.18−10.31% mIoU improvements show that group-wise attention benefits RUGD and RELLIS-3D across backbone choices.The ablation reports improvements regardless of the selected backbone.
  • Spatial Reduction: Higher spatial reduction decreases perception accuracy, but group-wise attention mitigates the resulting performance drop.The design trades spatial resolution for lower computational complexity.
  • Computational Cost: GANav has low memory requirements and the best inference time among listed transformer methods and most CNN-based methods.The broader complexity analysis considers parameters, GFLOPS, inference cost, and run-time.

E. Outdoor Navigation

GANav is evaluated on real-world navigation with Jackal and Husky robots against several planners. It improves navigation behavior in cluttered and unseen terrains, particularly by selecting smoother, more navigable surfaces.

  • Experimental Setup: GA-Nav is evaluated on Clearpath Jackal and Husky robots using RGB perception, a 3D lidar, and an elevation map.The comparison includes DWA, TERP, Segformer, and OCRNet-based planners.
  • Metrics: Success rate measures reaching the goal while avoiding relatively nontraversable regions and collisions.The navigation study also measures trajectory roughness, trajectory selection, and forbidden-region false positives.
  • Navigation Results: Segmentation-based planners perform well in Walkway and Trail scenarios where surface boundaries are clear.These scenarios resemble conditions with clearer separation between different surfaces.
  • Navigation Results: GA-Nav improves all three reported navigation metrics under cluttered and unseen Dirt road and Off-road conditions.The reported mechanism is more accurate segmentation, which increases use of the most navigable surface and improves trajectory selection.
  • Navigation Results: Elevation-based baseline planners cannot distinguish surface properties when different terrains have similar elevations, producing less navigable and rougher trajectories.Adding segmentation data enhances the baseline planner’s navigable trajectory selection.

V. CONCLUSIONS, LIMITATIONS, AND FUTURE WORK

The paper concludes that GANav provides efficient, accurate terrain segmentation for off-road navigation and demonstrates utility in real-world scenarios. The authors limit their analysis primarily to perception metrics and identify labeling consistency and broader navigation evaluation as open issues.

  • Conclusions: GANav improves accuracy and efficiency over state-of-the-art methods on complex unstructured terrain datasets and supports higher navigation success rates.The conclusion connects segmentation performance with real-world navigation use.
  • Limitations: The current analysis focuses mostly on perception metrics because the approach integrates with a learning-based planner rather than analyzing navigation comprehensively.The authors identify this as a limitation of the current work.
  • Conclusions: GA-Nav trajectories maximize travel along the smoothest available terrain while maintaining the highest success rate.The navigation comparison covers walkway, trail, dirt road, and off-road scenarios.
  • Future Work: Future work will evaluate more navigation metrics, adapt the method to other planning schemes, and incorporate multiple sensor inputs.These directions aim to support broader navigation evaluation and more robust navigation schemes.

APPENDIX

GANav groups terrains by texture and navigability so planning can account for robot-specific traversal capabilities. The grouping supports navigation comparisons and can be dynamically adapted to different robots and environments.

  • Terrain Grouping: Terrain characteristics such as texture help indicate navigability and distinguish surfaces relevant to robot navigation.The paper notes that similar textures and navigability can motivate grouping terrains into one category.
  • Terrain Grouping: Smooth surfaces such as concrete and asphalt are treated as navigable terrain.These surfaces correspond to the “flat” category in CityScapes and are generally traversable by mobile robots.
  • Terrain Grouping: Rough surfaces include gravel, grass, dirt, and sand, whose increased friction can affect traversal and planning.Planning may prioritize smooth navigable regions over rough navigable regions to reduce energy loss.
  • Terrain Grouping: Bumpy terrains such as small rocks and rock-bed can be assigned differently depending on robot capabilities.The grouping may be ignored or re-allocated for robots with stronger off-road capabilities.
  • Terrain Grouping: Water and bushes are categorized as forbidden regions, while trees and poles are treated as obstacles requiring detection for safe navigation.Background and non-navigable classes such as void, sky, and sign are included in a buffer group.
Loading 2103.04233v5…