Conferences and supporting programme
Innovations in Perception and Control for Autonomous Vehicle Operation
Autonomous driving is an apex challenge for sensing and computing. Many current market leaders use large, high-resolution sensors and array of processors to handle interpretation and decision-making on mountains of data every second. However, these solutions require a lot of power and generate a lot of heat - both of which are challenges for the vehicle itself. An alternative approach is to address the problem with smaller sensors and discrete, distributed compute elements providing energy efficiency operation. This paper describes the architecture of this new approach using a highly integrated programmable SoC platform that partitions functionality between an ARM-based quad-core processing complex and programmable logic for advanced perception via machine learning deployed at the edge, analytic decision-making and communications aspects for optimal operation of self-driving vehicles. The architectural description will include data flow details from acquisition to programmable SOC partitioning of the sensor data processing at both hardware and software functional levels. The unique sensor types, associated data rates, and timing will be described. The discussion on sensor and processing will cover the fusing of the data - sensor fusion, to provide a comprehensive understanding of the vehicle perception elements interactions and processing, including a new innovative modular software stack. An overview of the machine learning methodology utilized for the perception elements will be presented including both the training of the system and the deployment with associated comparison benchmarking in terms of performance, power, and responsiveness. Lastly, based on the above information, supplemented with V2X communications information, key elements of the command and control of the vehicle, will be presented to highlight the functional partitioning of the programmable SoC architecture for system operation. Real world scenarios will be presented, including showcasing vehicle operation independent of surrounding environment, demonstrating this innovative architecture along with benchmarks in terms of power, thermal, and algorithmic flexibility advantages over existing non-programmable SoC-based systems.
--- Date: 27.02.2018 Time: 3:00 PM - 3:30 PM Location: Conference Counter NCC Ost