Zhu, Minghao (The Ohio State University), Sidhu, Anmol (Transportation Research Center), Redmill, Keith (Ohio State University)

Enhancing Digital Speed Limit Detection in Work Zones: A Camera-Based Approach with Multi-Frame Processing

Scheduled for presentation during the Regular Session "Sensing, Vision, and Perception VI" (FrBT5), Friday, September 27, 2024, 15:10−15:30, Salon 13

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on October 3, 2024

Abstract

Twenty-eight percent of fatal crashes were speed-related traffic incidents in the US in 2021. Intelligent Speed Assistance (ISA) technologies show promise to reduce the severity and frequency of such accidents. Common sources of speed limit information used by ISA technologies are digital maps and perception systems. A digital map can be a straight forward and reliable source for speed limit information once roads are surveyed by map providers. However, due to the nontrivial resources and time needed to update a map, it can fail to reflect temporary speed limit changes such as variable speed limits (VSL) enforced in work zones. Therefore, a perception system is needed to detect the speed limits and supplement the digital map. Camera-based perception systems are widely used for traffic sign recognition (TSR) including speed limit detection. While speed limit signs work well in general for human drivers, data we have collected on public roads shows that speed limit detection can be challenging for vision-based TSR, and especially for digital speed limit (DSL) signs in work zones. In this paper, we describe the procedures used to investigate the issue of camera-based DSL detection and propose a solution to improve the detection performance. We construct a dataset that includes image frames captured by three different cameras and analyze detection performance based on different capture rates and camera models. Our proposed DSL sign recognition method utilizes multiple captures of a sign with a Segment Anything Model (SAM) for sign detection and a Swin-Transformer for feature extraction and achieves a 98.21% detection rate while requiring a frame rate of only 5 Hz.