UST 031

12 Lorenz Technology has launched its third generation of AI hardware and software for unmanned systems (writes Nick Flaherty). AI-Link uses an Nvidia processor with a dual Denver 64-bit core and four ARMv8 A57 cores alongside a 256-core Pascal graphics processing unit to handle machine learning algorithms on the unmanned platform, to detect incidents on pre-planned routes. It operates with UAVs or ground robots using the MavLink, DJI OSDK, or ROS operating systems. The 20 W AI-Link board measures 160 x 42 x 65 mm and weighs 426 g, Researchers in Israel have found that they can cause the autopilot on an autonomous vehicle to erroneously apply its brakes in response to ‘phantom’ images projected on a road or billboard (writes Nick Flaherty). The team, at the Ben-Gurion University of the Negev (BGU) Cyber Security Research Centre, shows that fully autonomous cars register depthless projections of objects – the phantoms – as real objects. They show how, without any special expertise, attackers can use a commercial UAV and an inexpensive image projector making it suitable for UAVs. The power consumption when idle is 7.5 W. An input voltage of 18-27 V is provided via an XT connector supporting up to 1.5 A, and the data connection is via USB 3.0, although Ethernet and serial UART connections are also possible. The board runs a deep neural network, which has been trained to identify objects and follow them, as well as specific functions such as the ability to detect containers at a shipping terminal and examine the condition of rescue ladders and fenders on port quays. Danish ferry operator DFDS has used AI-Link on a DJI M210 UAV for automated to manipulate the vehicle and potentially harm the driver or passengers. Relying only on an internal sensor creates a ‘validation gap’ with reality. The researchers are developing a neural network model that analyses the context, surface and reflected light of a detected object. This can detect phantoms with high accuracy said the team. To demonstrate the problem, the team projected traffic signs onto advertising billboards. These fooled the vehicle’s onboard systems when projected for just 125 ms. They also showed how fake lane markers projected on a road inspections of shipping and ports. Finding containers can require people to walk through a site, but a UAV using AI- Link can detect 300 trailers in 20 minutes, reducing the loading time from 6 minutes to 4.5 minutes per container An encrypted 4G data link to a cloud service called Hive provides live video with a latency of less than 500 ms, and all the data from the platform can be stored in the cloud to link to an existing data management system or for archiving. New plug-ins for the service can be developed in the cloud using training (called inference) and then downloaded to the AI-Link board. by a UAV could guide an autopilot into oncoming traffic. “This type of attack is not being taken into consideration by the automotive industry,” said Ben Nassi at the BGU’s Cyber Security Research Centre. “These are not bugs or poor coding errors but fundamental flaws in object detectors that are not trained to distinguish between real and fake objects and use feature matching to detect visual objects.” The problem occurs when the depthless objects projected on a road are considered real, even though the depth sensors can differentiate between 2D and 3D. AI system’s third version ‘Phantoms’ fool vehicles Artificial intelligence Driverless cars April/May 2020 | Unmanned Systems Technology The AI-Link uses an Nvidia processor on a Jetson board to implement machine learning algorithms on UAVs and UGVs