Low-viewpoint forest depth dataset for sparse rover swarms

Chaoyue Niu, Danesh Tarapore, Klaus-Peter Zauner

TL;DR

This paper introduces a novel, low-viewpoint forest depth dataset designed to train computer vision models for autonomous navigation in challenging outdoor environments. The dataset, comprising over 134,000 RGB-depth image pairs, addresses the critical need for realistic data that captures the unique perspective of small, ground-based robots, which is largely absent in existing datasets.

What Problem Does This Paper Address?

Autonomous navigation for robots, especially in complex environments like forests, hinges on accurately understanding the surrounding terrain. This requires precise depth estimation, which is crucial for tasks like path planning and obstacle avoidance. While monocular depth estimation techniques have advanced significantly, particularly with deep learning, existing public datasets often fall short. Many are recorded from high viewpoints (e.g., aerial or human-eye level), are limited in size, or focus on urban/indoor scenes. This leaves a gap for datasets that capture the low-to-the-ground perspective of small, ground-based robots, which is essential for developing robust navigation systems for applications like forest monitoring, search and rescue, or environmental protection using robot swarms.

Methods

The authors developed a custom data collection rig to capture synchronized RGB and depth data from a low viewpoint. This rig featured an Intel D435i depth camera, two rotary encoders for precise odometry, and a GPS receiver. The system was manually pushed through forest environments to simulate the movement of a ground rover. Data was collected under various weather and lighting conditions to ensure diversity. Post-collection, the raw data was processed to synchronize sensor streams (RGB, depth, GPS, IMU, encoders) and select high-quality image-depth pairs, resulting in approximately 9700 aligned frames. The paper also details methods for assessing data quality, including "fill rate" (percentage of valid depth pixels) and depth accuracy against ground truth measurements from a laser distance meter.

Datasets

The primary contribution is the newly introduced forest depth dataset. This dataset contains:

  • Over 134,000 RGB/depth image pairs.
  • Approximately 9700 aligned RGB and depth image pairs with synchronized metadata (GPS, IMU, rotary encoder data).
  • Data collected from a low viewpoint, mimicking small ground robots.
  • Diverse forest scenes with various vegetation, lighting, and weather conditions.

The dataset is made publicly available under a Creative Commons Attribution 4.0 International Public License.

Key Insights and Ideas

  • The "low-viewpoint" challenge: Existing depth datasets are often not representative of the perspective of small, ground-based robots, which is a critical factor for their navigation capabilities.
  • Forest complexity: Forests present unique challenges for depth estimation due to dense foliage, occlusions, and varying terrain, requiring specialized datasets.
  • Data diversity is key: Collecting data across different times of day, weather conditions, and lighting is essential for training robust models that can generalize well.
  • Quality assessment is crucial: The paper emphasizes the importance of evaluating depth data quality through metrics like fill rate and accuracy, demonstrating that their collected data maintains a reasonably high fill rate (mean 0.84) even under varying velocities and luminosities.

Key Results

  • A dataset of over 134,000 RGB-depth image pairs was collected from a low viewpoint in a forest environment.
  • A subset of ~9700 aligned image-depth pairs with synchronized metadata is provided for research.
  • Depth accuracy was evaluated, showing a mean error of less than 4% across sampled points, with the highest deviation at the furthest point.
  • The fill rate of depth images was found to be consistently high (mean 0.84 ± 0.11 SD) and largely unaffected by recording velocity or scene luminosity within the tested ranges.
  • The dataset is intended to facilitate the training of deep neural networks for monocular depth estimation and other computer vision tasks relevant to autonomous forest navigation.

Limitations and Future Work

  • Dataset Size for Deep Learning: While the dataset is substantial, the authors acknowledge that even larger datasets might be needed for training state-of-the-art deep neural networks, especially for complex tasks.
  • Manual Data Collection: The data was collected by manually pushing the rig. While this simulates a ground robot's perspective, it doesn't fully replicate the dynamic motion and control of an actual autonomous robot.
  • Future Work: The authors plan to use this dataset to develop depth prediction models for low-capability embedded computers, aiming to enable sparse robot swarms. They also suggest the metadata (especially rotary encoder data) could be used to train autonomous guidance systems.

Summary updated

BibTeX
@inproceedings{niu2020low,
  author = {Chaoyue Niu and Danesh Tarapore and Klaus-Peter Zauner},
  title = {{Low-viewpoint forest depth dataset for sparse rover swarms}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2020},
  doi = {10.1109/iros45743.2020.9341435},
}
Figure