Overview / Usage

Autonomous and intelligent flight under the canopy of densely forested areas remains an unaddressed challenge. It requires a UAV to identify an optimal flight route within an unseen environment based solely on visual perception.

Changes in the domain such as weather and illumination directly affect any algorithmic ability to correctly define flight direction. We address this issue by training a deep neural network (DNN) with a single forward facing camera view. This image is then cropped into {left, right, forward} partitions which can be labelled for trail presence/absence.

Methodology / Approach

DNN training uses a gradient descent optimizer, random weight initialisation with zero node biases and is performed over 90 epochs with a learning rate decay rate of 0.95 per epoch. For both training and testing, we used the high-resolution (752 x 480) IDSIA dataset and a low resolution (106 x 240) Urperth Burn (UB) dataset, gathered locally. After training the model was ported to the Intel Movidius Neural Compute SDK (NC SDK) were we tested the accuracy and inference time.

Technologies Used

Training is performed in the Intel AI DevCloud using the Intel Optimization for Tensorflow and Intel Distribution for Python. Next, data gathered by the Intel Ready to Fly Drone will be tested on the Intel Movidius Neural Compute SDK, during flight.

Repository

https://github.com/brunapearson/movidius

Other links

• Maciel-Pearson B.G., Carbonneau P., Breckon T.P. (2018) Extending Deep Neural Network Trail Navigation for Unmanned Aerial Vehicle Operation Within the Forest Canopy. In: Giuliani M., Assaf T., Giannaccini M. (eds) Towards Autonomous Robotic Systems. TAROS 2018. Lecture Notes in Computer Science, vol 10965. Springer, Cham.