This work provides a method for real-time visual navigation for ground robots in unstructured environments. The algorithm uses a convolutional network to propose a set of trajectories and costs for efficient and collision-free movement from a single webcam. The models in this project are trained to specifically target forests. For more details, see the paper.
Figure: Top three trajectories (green, yellow, red) predicted from a forest scene.
Setup: Clone the repository and follow the instructions in the headers of each of the files listed below.
- Generate an NDVI prediction model from
ndvi-model/model.py - Label datasets using
regional-cost-model/label.py - Train the model with
regional-cost-model/model.py - Convert the model to TFLite with
tflite-model/tflite.py - Run the TFLite model online with
tflite-model/tflite_standalone.py
Notes: Prior to step 5, it may be necessary to resolve the order of the model's output heads. See Appendix D of the paper and tflite-model/reorder.py. For further information on any of the steps, see the paper and included comments in the code.
- TensorFlow 2.x
- OpenCV 4.x
- scikit-learn 1.x
- TFLite 2.x
- SciPy 1.x
- DroneKit 2.x
- pymavlink 2.x
Other miscellaneous libraries (e.g. matplotlib) will be needed to run the code but are not required.
- Hardware: Raspberry Pi 4B (TBC GiB)
- Traversal Cost Labelling Parameters: (γ, α, β) = (0, 1, 0.1)
- Model Parameters:
- Optimizer: Adam
- Batch Size: 32
- Loss: MAE
- Epochs: 100
- Train-Test Split: 80-20
- Number of Instances: 9700
This work is all the original work of the author and is distributed under the GNU General Public License v3.0. For more details, see the LICENSE file.