10.2 - PARIS - Parallelisation Architecture for Real-time Image Data Exploitation and Sensor Data Fusion
- ettc2018 - European Test and Telemetry Conference
2018-06-26 - 2018-06-28
- 10. Imaging & Video
- S. Blokzyl, M. Nagler, R. Schmidt, W. Hardt - Computer Science Department, Chemnitz University of Technology,Chemnitz (Germany)
- 217 - 223
The development of visual recognition systems for highly automated, mobile systems is driven by matured electro-optical sensors and application-specific, optimised computational components. Applying good-quality, high-resolution imagery exploited by powerful embedded processors, current computer vision systems provide a detailed representation of system’s operational environment and contribute significantly to various perception tasks. However, the exploitation of high-resolution image data is challenging on space-, weight-, and energy-limited mobile systems, especially when respecting realtime and reliability requirements of safety critical functions.
Considering these conditions, this paper introduces a generic parallelisation architecture for real-time image processing using reconfigurable integrated circuits. The architecture supports data and task parallelisation strategies utilising the specific parallelisation capabilities of Field Programmable Gate Arrays. It provides space-, weight-, and energy-efficient parallelisation for image exploitation and information fusion.
The device- and interface-independent architecture maximises parallelism and flexibility of complex image processing applications aboard mobile systems. The modular structure of the proposed architecture enables hardware-acceleration for high-resolution sensor data exploitation, minimises processing latency, and improves the quality of the overall detection result by using multivariate detection methods.
The generic parallelisation architecture will be used for multi-copter-based high voltage transmission line inspection and vision-based localisation of an unmanned aerial vehicle during final approach phase.