Efficient Vision Transformers for Autonomous Off-Road Perception Systems  

作　　者：Max H. Faykus III Adam Pickeral Ethan Marquez Melissa C. Smith Jon C. Calhoun Max H. Faykus III;Adam Pickeral;Ethan Marquez;Melissa C. Smith;Jon C. Calhoun(Holcombe Department of Electrical and Computer Engineering, Clemson University, Clemson, SC, USA)

机构地区：[1]Holcombe Department of Electrical and Computer Engineering, Clemson University, Clemson, SC, USA

出　　处：《Journal of Computer and Communications》2024年第9期188-207,共20页电脑和通信（英文）

摘　　要：The development of autonomous vehicles has become one of the greatest research endeavors in recent years. These vehicles rely on many complex systems working in tandem to make decisions. For practical use and safety reasons, these systems must not only be accurate, but also quickly detect changes in the surrounding environment. In autonomous vehicle research, the environment perception system is one of the key components of development. Environment perception systems allow the vehicle to understand its surroundings. This is done by using cameras, light detection and ranging (LiDAR), with other sensor systems and modalities. Deep learning computer vision algorithms have been shown to be the strongest tool for translating camera data into accurate and safe traversability decisions regarding the environment surrounding a vehicle. In order for a vehicle to safely traverse an area in real time, these computer vision algorithms must be accurate and have low latency. While much research has studied autonomous driving for traversing well-structured urban environments, limited research exists evaluating perception system improvements in off-road settings. This research aims to investigate the adaptability of several existing deep-learning architectures for semantic segmentation in off-road environments. Previous studies of two Convolutional Neural Network (CNN) architectures are included for comparison with new evaluation of Vision Transformer (ViT) architectures for semantic segmentation. Our results demonstrate viability of ViT architectures for off-road perception systems, having a strong segmentation accuracy, lower inference speed and memory footprint compared to previous results with CNN architectures.The development of autonomous vehicles has become one of the greatest research endeavors in recent years. These vehicles rely on many complex systems working in tandem to make decisions. For practical use and safety reasons, these systems must not only be accurate, but also quickly detect changes in the surrounding environment. In autonomous vehicle research, the environment perception system is one of the key components of development. Environment perception systems allow the vehicle to understand its surroundings. This is done by using cameras, light detection and ranging (LiDAR), with other sensor systems and modalities. Deep learning computer vision algorithms have been shown to be the strongest tool for translating camera data into accurate and safe traversability decisions regarding the environment surrounding a vehicle. In order for a vehicle to safely traverse an area in real time, these computer vision algorithms must be accurate and have low latency. While much research has studied autonomous driving for traversing well-structured urban environments, limited research exists evaluating perception system improvements in off-road settings. This research aims to investigate the adaptability of several existing deep-learning architectures for semantic segmentation in off-road environments. Previous studies of two Convolutional Neural Network (CNN) architectures are included for comparison with new evaluation of Vision Transformer (ViT) architectures for semantic segmentation. Our results demonstrate viability of ViT architectures for off-road perception systems, having a strong segmentation accuracy, lower inference speed and memory footprint compared to previous results with CNN architectures.

关 键 词：Semantic Segmentation Off-Road Vision TRANSFORMERS CNNS Autonomous Driving 

分 类 号：TN9[电子电信—信息与通信工程]