Deep learning-based caching optimization for VR 360° videos in vehicular edge computing

The rapid advancement of Virtual Reality (VR) technologies has revolutionized immersive 360° video streaming, especially in Vehicular Edge Computing (VEC) environments. However, delivering seamless VR experiences is challenging due to high bandwidth requirements, ultra-low latency, and dynamic user viewports, which are exacerbated by vehicular mobility. Traditional caching strategies, such as Least Frequently Used (LFU) and Least Recently Used (LRU), and even recent viewport-aware methods struggle to adapt to these spatio-temporal dynamics. To overcome these limitations, this paper proposes DeepEdge360, a deep learning-based framework for optimizing the caching of 360° videos in VEC. Our solution integrates adaptive tile-based segmentation, viewport-aware prioritization, and proactive prefetching to dynamically align caching decisions with user behavior and vehicular mobility. In particular, the DeepEdge360 is threefold: First, a spatio-temporal tile segmentation and request mechanism is introduced, utilizing Long Short-Term Memory (LSTM)-based popularity prediction to anticipate user viewing patterns. Second, a proactive caching strategy is implemented to optimize storage space in vehicles and Roadside Units (RSUs), adapting it to user behavior and vehicular mobility. Third, a Deep Q-Network (DQN)-based eviction strategy is proposed to intelligently manage cache replacements, thus balancing cache hit rates, latency, and bandwidth utilization. The proposed DeepEdge360 framework significantly outperforms traditional and state-of-the-art caching schemes, achieving an 82% cache hit rate, 45ms end-to-end latency, and 76% bandwidth utilization. The results validate our framework's effectiveness in supporting high-quality VR streaming in dynamic vehicular networks, while its modular design ensures practical deployability in edge-assisted environments.