An Objective Evaluation of Edge Devices Optimized with Hardware Processors Running a Real-Time Machine Vision System Targeting Autonomous Cars

Abstract

To navigate safely, autonomous vehicles require a full understanding of their environment by not only detecting what objects are surrounding the vehicle but also by knowing how far they are. In addition, such systems have to process the information perceived in real-time in order to take action promptly and avoid accidents. The purpose of this Thesis is to design and build an embedded parallel vision system, targeting moving vehicles that provides a real-time system able to identity common objects in driving scenes, namely vehicles, pedestrians, bicycles, and traffic signs; and estimate their distance. Moreover, the proposed system is deployed in three different NVIDIA edge devices, and their inference throughput, speed and power consumption is evaluated and compared. Finally, optimizations are been performed on both software and hardware in order to accelerate the overall efficiency of the system.

Date of Award	14 Dec 2023
Original language	American English
Supervisor	MAHMOUD Meribout (Supervisor)