ADR-Net, a deep convolutional neural network model for detection,
works real-time with our MAV camera setting on NVIDIA TX1 embedded board.
ADR-Net is based on AlexNet and SSD architecture but much faster.
The base network used was the AlexNet. As our purpose was a simple gate detection,
we reduced the number of unnecessary high-level feature layers.
In addition, we considered the drone-racing arena,
where the distance between the gates did not exceed 5 m.
Therefore, a regression operation to find a small object could be avoided.
We finally achieved the desired speed of about 30 fps, but owing to the layer deletion,
the average precision was reduced by 0.13 compared to the original SSD model.
To compensate for the accuracy reduction, we tuned the parameters of the network.