FPGA acceleration of computationally intensive embedded applications
The technology that will be provided by the research group refers to specialized hardware (HW) accelerators for computationally intensive embedded applications. The proposed accelerators can be used in various application domains, e.g., in advanced driver assistance systems for autonomous vehicles enabling real-time processing and/or autonomous ships for maritime mobility. The functionality of our currently developed accelerators has been validated in real HW boards via extensive testing and by using realistic datasets. Our experimental results demonstrate substantial performance gains in terms of speed compared to conventional implementations on central processing units (CPUs)
In the current project we propose a pool of HW accelerators IPs including general-purpose convolutional neural networks (CNNs) for image processing. The implementation of the accelerators is based on specialized high-performance chips namely Field Programmable Gate Arrays (FPGAs).
The most prominent advantages of FPGAs are: i) the deployment of highly parallel designs such as neural networks with impressive throughput rates, ii) the reduced power consumption, and iii) the reconfiguration/reprogramability which provides increased flexibility and re-usability of the chip for a broad range of applications and operations.
Stage of development
We have already implemented several HW accelerators on FPGAs including convolutional neural networks (CNNs) for image processing (TRL3). The functionality of our currently developed accelerators has been validated in real HW boards via extensive testing and by using realistic datasets. Our experimental results demonstrate substantial performance gains in the order of 10-100x in terms of speed compared to conventional implementations on CPUs.
Challenge and needs
One of the most computationally intensive operations in advanced driver assistance systems is related to image processing for object detection, traffic detection, etc. Modern systems are expected to employ combination of both camera and light detection and ranging (LiDAR) sensors posing a great computational challenge due to the high amount of data that needs to get processed at real-time. Currently, there is a lack of implementations for LiDAR-based systems involving operations such feature detection and feature extraction. The solution that we propose is the implementation of a HW accelerator on FPGA for real-time LiDAR data processing. Specifically, we seek for the implementation of a HW efficient convolutional neural network (CNN) for image processing supporting the operation of feature detection and extraction on the data captured by LiDAR sensor. The implementation of the accelerator will be based on currently developed and functional parametric IPs which will be customized accordingly to address the requirements of the respective system. The training and the evaluation of the implemented CNN will be based on currently available realistic datasets such as the Ford and KITI benchmarks.
A prior art search is scheduled to be done.
Potential markets and targets
FPGA-based systems for processing of LiDAR data can be applied to many modern embedded systems and a lot of companies could benefit from such development and exploitation. Markets range from automotive applications to maritime industry and autonomous shipping. According to recent Mackinsey report there is a tremendous growth for the following/potential markets:
Partners (SMEs) which could have business interest to the outcome of the project and also to the co-development of the proposed solution could be: