Design and Evaluation of a Real-Time Semantic Segmentation System for Autonomous Driving

Autonomous driving represents a revolutionary technology poised to reshape the automotive industry, offering the promise of safer, more efficient, and environment friendly vehicles. Real-time semantic segmentation is considered as a fundamental component of autonomous driving systems, playing a pivotal role in scene comprehension and decision-making. In this paper, an attempt is made to design, analyze and evaluate real-time semantic segmentation algorithms for autonomous driving. The proposed algorithms and architectures are initially tested on a laptop with GPU and later implemented on two distinct embedded platforms: Raspberry Pi 4 (8GB RAM) board and Xilinx Zynq 7000 FPGA-based PYNQ (Python Productivity for Zynq) Z2 development board. The performance evaluation of proposed real-time semantic segmentation system using various hardware and architectures are reported in this paper. The maximum Mean Intersection of Union (mIoU) obtained from proposed system is 0.8315 and the best refresh rate achieved is 18.32. The proposed system had few limitations on the PYNQ board, whereas the Raspberry PI board gave satisfactory results. The proposed work can be easily extended to other hardware and offers insights that can be leveraged to optimize autonomous driving systems for edge devices.