Abstract— Real-time collision prediction is critical for the safety and reliability of autonomous driving systems. However, accurately forecasting collisions remains challenging due to the complexity of dynamic driving environments, noisy sensor data, scarcity of rare collision events, limited adaptability to new scenarios, and strict real-time constraints. Existing collision prediction models typically process spatial and temporal data separately, struggle to filter sensor noise effectively, and require extensive retraining for new conditions, hindering practical real-time deployment. This paper introduces Hybrid Deep Collision Prediction Network (HDC-Net), a novel unified framework designed to overcome these challenges. HDC-Net integrates a dilated convolutional neural network (CNN) with a dual-branch recurrent neural network (RNN) comprising Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) modules to jointly capture spatial context and multi-scale temporal dynamics. A self-attention mechanism is embedded to filter sensor noise and highlight essential collision indicators. Additionally, HDC-Net employs Model-Agnostic Meta-Learning (MAML) to facilitate rapid model adaptation to new driving conditions and leverages generative adversarial networks (GAN) for synthesizing realistic collision scenarios, addressing data scarcity. Computational optimizations, including hierarchical attention pooling and kernel fusion, ensure the model’s real-time operability. The performance of HDC-Net was evaluated on the DeepAccident dataset using a rigorous 4-fold cross-validation. Results demonstrate that HDC-Net achieves a collision prediction accuracy of 89.3%, a time-to-collision error of 0.42 seconds, a trajectory deviation of 0.17 meters, and an inference speed of 18.4 ms per frame. Compared to state-of-the-art baselines, HDC-Net significantly improves prediction accuracy by approximately 4.6% while maintaining efficient real-time performance.