Taking Flight with Tinker Edge T
Tinker Edge T is a series of small and powerful single board computers (SBC) featuring a neural network accelerator processor (NPU) that can run machine learning (ML) inference in a power-efficient manner and bring artificial intelligence processing from the cloud to a local device. With an onboard Google TPU, Tinker Edge T can perform AI model training under Google TensorFlow Lite and run AI inference.
ASUS is sponsoring RWTH Aachen University's Aachen Drone Development Initiative (ADDI) in Germany, which is part of the university’s unmanned aerial vehicle (UAV) project. Students in the ADDI will compete in the International Micro Air Vehicle Conference (IMAV) event being held in Mexico in 2021. The UAV the students will enter in the competition using Tinker Edge T as its onboard computer. Directly attached to the UAV, Tinker Edge T will be used to analyze images and videos in real time, perform mapping tasks and detect drop zones as well as handling the communication with the flight controller and the ground station.
IMAV 2021 is set in a post disaster scenario, for example earth quacks, with the main goals being search and rescue as well as delivery of necessary aid. One goal is for the UAV to detect when the ground is safe to land and drop packages on. A neural network is deployed to reliably determine this from camera footage of an onboard camera connected to the tinker edge.
Built upon the MobileNetV2 object detection network as a basis, they’ve opted for a convolutional auto-encoder with skip-connections performing image segmentation. Relying on a proven image detection network makes sure all kinds of objects are detected, whereas the custom-trained upscaling branch of the network determines if they are considered obstructive.
While the network is trained using cloud services on computer-generated samples, the inference has to happen onboard the drone at a reasonable speed, as multiple drones searching simultaneously prohibit a centralized solution. Using the Tensorflow Lite framework to infer the network on the embedded TPU of the Tinker Edge T leaves the CPU free for post-processing tasks and time-critical control loops, allowing for a painless integration.