TimeSync: GAN-Driven Temporal Feature Synchronization for Robust Collaborative Perception in Autonomous Driving
Yalong Li, Zhaoyang Du, Rui Yin, Wugedele Bao, Ha Si, Celimuge Wu · 2024
Cooperative perception is crucial for enhancing the safety and efficiency of autonomous driving systems.However, temporal misalignments caused by varying communication delays between vehicles pose significant challenges to accurate feature fusion.To address this issue, we present TimeSync, a novel adversarial learning-based framework for temporal feature alignment in cooperative perception.Specifically, we first propose a Time-Aware Residual Network generator with a Dynamic Delay Embedding module, coupled with a Temporal PatchGAN Discriminator, which effectively handles varying communication delays.Moreover, TimeSync introduces a comprehensive multi-component loss function that enhances alignment accuracy across different time steps and vehicles.Additionally, we develop an efficient multi-vehicle feature fusion approach that preserves spatial and temporal coherence.Extensive experiments on large-scale autonomous driving datasets demonstrate that TimeSync significantly outperforms existing approaches in terms of alignment accuracy and adaptability to diverse driving scenarios.