Unmanned Systems Technology 021 | Robot Aviation FX450 l Imaging Sensors focus l UAVs Insight l Liquid-Piston X-Mini l Riptide l Eurosatory 2018 show report l Zipline l Electric Motors focus l ASTS show report

10 Platform one August/September 2018 | Unmanned Systems Technology FLIR Systems has developed a data set for machine learning that uses thermal images (writes Nick Flaherty). It is aimed at researchers, developers and makers of self-driving vehicles The data set contains more than 10,000 annotated thermal images of day and night-time scenarios for training convolution neural networks (CNNs), and is available for free. It is the first data set to include annotations for cars, other vehicles, people, bicycles and dogs. The starter thermal data set enables developers to begin testing and evolving CNNs with the FLIR Automotive Development Kit. The data set speeds up the evaluation of thermal sensors on next-generation algorithms. When combined with visible light cameras, Lidar and radar, and thermal sensor data paired with machine learning, it provides a more comprehensive system for identifying and classifying roadway objects, especially pedestrians and other living things, providing redundancy alongside electro- optic cameras.  “It provides a critical opportunity for the automotive community to expand the data set to make self-driving cars more capable in various conditions,” said Frank Pennisi at FLIR. “With the potential for millions of autonomous-enabled vehicles, thermal sensor costs will decrease significantly, which will encourage widescale adoption and ultimately enable safer autonomous vehicles.” Imaging Free thermal images data Better machine learning algorithms rather than more sensors are the key to more effective driverless cars, says a UK start- up (writes Nick Flaherty). Researchers at Wayve are developing end-to-end machine learning algorithms for autonomous vehicles using a new approach that does not need a lot of training. “The missing piece of the self-driving puzzle is intelligent algorithms, not more sensors, rules and maps,” said Amar Shah at Wayve. “Humans have a fascinating ability to perform complex tasks in the real world, because our brains allow us to learn quickly and transfer knowledge across our many experiences. We want to give our vehicles better brains, not more hardware.” Wayve is testing its algorithms on public UK roads using a prototype vehicle. This is a modified Renault Twizy two-seater that has a range of 100 km on a single battery charge. It uses a single monocular forward- facing video camera mounted in the centre of the roof at the front of the vehicle. It also has retrofitted electric motors to actuate the brakes and steering, and electronically emulate the throttle position to regulate torque to the wheels. All the control and computation is handled by a single Nvidia Drive PX2 computer. This runs Wayve’s software that is based on an approach called reinforcement learning, which doesn’t need the vehicle to build a complex model of its environment. This strategy, called deep deterministic policy gradients, enabled the test car to learn how to follow a lane to 96% accuracy in 20 minutes, and 98% after two hours. The vehicle continues learning to improve the accuracy of the reinforcement learning algorithm. “Building a self-driving system that can safely drive on roads is too difficult to hand-engineer,” said Alex Kendall, also at Wayve. “We’re betting on technology such as reinforcement and imitation learning, and think machine learning will provide the breakthrough to deliver autonomous vehicles for everyone, everywhere.” Autonomous vehicles ‘Better brains’ for cars Wayve is testing its machine learning algorithms on UK roads using a modified Renault Twizy