7 Platform one Uncrewed Systems Technology | October/November 2023 Two UK projects are aiming to improve the sensor and data models used for simulations to develop autonomous systems (writes Nick Flaherty). Both projects include rfPro, which makes simulation hardware and software systems to test equipment as part of systems development. The first project, SIM4CamSense, aims to improve the sensor models in the simulations, and is working in collaboration with the UK’s Compound Semiconductor Applications Catapult and National Physical Laboratory to produce models of the physical camera, radar and Lidar sensors. The second project, DeepSafe, is developing a machine learning system to generate test data to train autonomous vehicles to handle the rare, unexpected edge cases that vehicles must be prepared to encounter on the road. The SIM4CamSense project builds on rfPro’s work with Sony to characterise the Sony image sensor to create a more accurate model that can be used in a simulator. “This can be the chip or full sensors, and our job is to improve the models to correlate to the measurement data from the lab in controlled field trials for all the sensor modalities, the radar, Lidar and the camera,” said Matt Daley, operations director for rfPro. The company aims to include the sensor manufacturers in the project’s steering committee so that the models can be used across industry. “We already have the Sony CMOS chip model in rfPRO but we will be actively encouraging chip makers, image signal processor providers, Lidar and radar companies to participate,” he said. “The best thing we can do is encourage them to provide a black box model. “Sony describes it as the ‘inside out’ model for us to deliver our data. We want that level of detail from the radar and Lidar providers for physical-based models for all the important sensors and chipsets.” This will be used to improve the modelling of how sensors detect particular materials, which is different for image sensors, radar and Lidar. “We have a physical material ID in rfPro so we no longer try to make up an electromagnetic reflectance value; we let the sensor model hit every pixel to create a material ID,” Daley said. “The objective is to integrate the whole process of physical material IDs into the simulation assets.” The DeepSafe project aims to use the models to allow developers to test the sensor systems more thoroughly in simulations. This is led by dRISK.ai with Imperial College London, Claytex and rFpro to develop the simulationbased training data. “dRISK are leading the project on developing the right training data,” Daley said. “They have a knowledge graph with a database of edge cases, with on-road data from sensors, and this is about improving the models and software to simulate as many of the edge cases as possible. Part of the project is to expand the capability of the simulation assets.” One way for a simulator to produce the synthetic data is to have procedures to generate many different use cases to cover the edge conditions, rather than having to store all the data in a huge database. “This will be as procedural as possible, but we are also working to improve the animation APIs for more motion capture input,” Daley said. Simulation Better models projects Developing sensor models for synthetic data
RkJQdWJsaXNoZXIy MjI2Mzk4