USE Network launch I UAV Works VALAQ l Cable harnesses l USVs insight l Xponential 2020 update l MARIN AUV l Suter Industries TOA 288 l Vitirover l AI systems l Vtrus ABI

83 The processing requirements for these frameworks is growing exponentially, so engineers are now looking at how these networks can be optimised to fit into smaller, custom devices rather than general-purpose GPUs. That requires a more complex relationship between the hardware design, the neural network framework and software tools such as compilers and optimisers. There are also new types of neural network that can reduce the power consumption by mimicking biological neural networks. The neuromorphic networks, also called spiking neural networks (SNNs), can be event-driven, responding to changes in a data signal rather than having to look over a whole frame. New chips using SNNs are being used for radar systems on UAVs and for implementing complex path-planning algorithms. Performance requirements One of the ways these networks are being implemented in inference engines is on small boards. An alternative is a custom module that is specifically designed to run the popular Tensorflow framework. This Edge Tensorflow Processing Unit (TPU) runs 4 TOPS (Tera operations per second) with a power consumption of 2 W, allowing it to be reduced in size to fit into UAVs and AGVs (automated guided vehicles). The board can run a Tensorflow framework such as MobileNet v2 that has been trained on a desktop or data centre system at 400 frame/s. The alternative Jetson board uses dual custom-designed 64-bit ARM and four ARMv8 A57 cores alongside a 256-core Pascal GPU to handle machine learning algorithms on UAVs or ground robots using the MAVLink, DJI OSDK, or ROS operating systems. The board runs a DNN that has been trained to identify objects and follow them. That gives it the ability to detect containers at a shipping terminal, as well as examine the condition of rescue ladders and fenders on port quays. For example, a Danish ferry operator has used the board on a UAV for automated inspections of shipping and ports. The AI board can detect 300 trailers in 20 minutes, reducing the loading time from 6 to 4.5 minutes per container. FPGAs are also a key tool in the implementation of inference engines. The technology is already being used in camera-based safety systems for processing video and extending AI systems | Focus Unmanned Systems Technology | June/July 2020 Requirements for machine learning are growing exponentially in terms of Tera-operations per second (Courtesy of Rambus) An Edge Tensorflow Processing Unit has been mounted in a module on a small board suitable for UAVs (Courtesy of Murata) A 20 W board measuring 160 x 42 x 65 mm weighs 426 g, making it suitable for UAVs. The power consumption when idle is 7.5 W (Courtesy of Nvidia)