Design and Implementation of Artificial Intelligence Models Using Deep Neural Networks on Reconfigurable VLSI Systems for Autonomous Driving

Abstract

Autonomous vehicles use object detection in real-time, which is an important aspect of navigation and decision-making systems. Nevertheless, conventional computing architectures like CPUs and GPUs are usually inadequate to support the demanding latency, power consumption and real-time demands within embedded automotive systems. This paper gives details of a generic design and implementation of object detection models using deep neural networks on reconfigurable Very Large Scale Integration (VLSI) systems including Field-Programmable Gate Arrays (FPGAs). The quantized and compressed architectures of DNN are shown as combining a system-level co-design approach with an FPGA platform by means of optimized-mapping of hardware and parallel dataflow design. The framework has been proposed as based on low-latency, high-throughput, energy-efficient inference, which can be brought to the edge in when safety is required. The process of simulation and hardware synthesis entails MATLAB, Simulink, HDL Coder, and Xilinx Vivado, with experimental analysis being carried out on real datasets, such as KITTI or BDD100K. Experiments show that indeed there is a huge gain in the rate of number of inferences per second and resource consumption as well as power generations when compared to typical CPU/GPU deployment. The results support the conclusion on the usefulness of reconfigurable VLSI platforms as an alternative hardware solution to building autonomous driving systems by AI in the future.