# StereoVoxelNet: Real-Time Obstacle Detection Based on Occupancy Voxels from a Stereo Camera Using Deep Neural Networks

**Authors:** Hongyu Li, Zhengang Li, Neset Unver Akmandor, Huaizu Jiang, Yanzhi Wang, Taskin Padir  
**arXiv:** [2209.08459](https://arxiv.org/abs/2209.08459)  
**Date:** 2022-09-18  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2209.08459v1/

## Key points

- Paper title: "StereoVoxelNet: Real-Time Obstacle Detection Based on Occupancy Voxels from a Stereo Camera Using Deep Neural Networks"
- Addresses the safety-critical problem of obstacle detection in robot navigation
- Proposes a computationally efficient method for obstacle detection using deep neural networks and volumetric representations
- Introduces a coarse-to-fine pruning technique to eliminate computation for safety-irrelevant spaces
- Achieves accurate obstacle detection within a range of 32 meters
- Outperforms state-of-the-art stereo models in terms of Intersection over Union (IoU) and Chamfer Distance (CD) scores while requiring only 2% of their computation cost
- Validates robustness and real-world feasibility through autonomous navigation experiments with a real robot
- Provides a 1.36-hour stereo dataset, code, and additional visualizations for further research and development

## Summary

The paper titled "StereoVoxelNet: Real-Time Obstacle Detection Based on Occupancy Voxels from a Stereo Camera Using Deep Neural Networks" addresses the safety-critical problem of obstacle detection in robot navigation. The authors propose a computationally efficient method that directly detects occupancy from stereo images using a deep neural network. Instead of learning point cloud correspondence from stereo data, the proposed approach extracts compact obstacle distribution based on volumetric representations. A coarse-to-fine pruning technique is introduced to eliminate computation for safety-irrelevant spaces using octrees generated by the decoder, resulting in real-time performance on an onboard computer (NVIDIA Jetson TX2). This method achieves accurate obstacle detection within a range of 32 meters and outperforms state-of-the-art stereo models in terms of Intersection over Union (IoU) and Chamfer Distance (CD) scores while requiring only 2% of their computation cost. To validate its robustness and real-world feasibility, autonomous navigation experiments were conducted with a real robot. A 1.36-hour stereo dataset was collected with a Jackal robot to fine-tune the model; this dataset is made available by the authors along with code and additional visualizations. In summary, this work contributes to bridging the gap between stereo systems in robot perception and state-of-the art stereo models in computer vision by leveraging deep neural networks and volumetric representations for efficient real-time obstacle detection in robot navigation tasks.