Dual-Attention-Guided Traffic Event Video Classification Network
Chunsheng Liu, Penghui Hao, Faliang Chang, Jun Zhou, Zehao Liu · 2024
A Dual-Attention-Guided Time-Shifted Convolutional Traffic Event Video Classification Network (DAG-Net) is proposed to classify events in traffic videos, addressing the challenges posed by complex and changing traffic environments and varying relative camera positions. Dual Attention Based Feature Enhancement Module (DAE-Module) is designed to focus on the main objects, such as cars and pedestrians, thus addressing the challenges of diverse traffic backgrounds and rapidly changing environments. Additionally, we introduce the Weighted Time Shift Module (WTS-Module) to handle rapidly changing traffic video environments by improving the network's capability to capture temporal relationships between consecutive frames. We built a Traffic Event Video Classification and Recognition (TEVCR) dataset, and experiments on the TEVCR dataset demonstrated the good performance of this method.