Wang, Xiaoqin (Chongqing University), Yin, Yunfei (Chongqing University), Zhang, Caizhi (Chongqing University), Bao, Xianjian (Maharishi University of Management)

MLM-Former: Enhanced Video Instance Lane Detection Via Spatio-Temporal Memory and High-Level Lane Features

Scheduled for presentation during the Poster Session "Transformer networks" (WeBT13), Wednesday, September 25, 2024, 14:30−16:30, Foyer

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 December 26, 2024

Abstract

Accelerated by the electric vehicle market’s growth, research in autonomous driving emphasizes the critical role of lane detection in the development and pricing strategies of the electric vehicle market. Currently, lane detection methods primarily focus on processing static images, neglecting dynamic changes, thus limiting their application in complex environments. In the realm of lane detection research, attention needs to be directed not only towards dynamic features but also towards lane characteristics. To tackle these challenges, a new video lane detection network has been proposed. This network utilizes a CNN-based spatio-temporal memory network to input extracted low-level spatio-temporal lane memory features into the High-Level Lane Transformer (HLaneformer) module, effectively learning high-level lane features. Simultaneously, a lane memory is constructed to aggregate the features of target lanes. This enables a deep exploration of the low-level lane features predicted by the CNN network, effectively compensating for the limitations of CNN in handling thin and distant lane shapes. Experimental results demonstrate that MLM-Former outperforms existing video lane detection methods on three public datasets, achieving state-of-the-art performance.