Abstract

Autonomous vehicles require robust early collision prediction to ensure safety, yet sudden maneuvers and changing environments make accurate forecasting challenging. Existing methods often process spatial and temporal data separately, limiting predictive reliability. We propose the Hybrid Deep Collision Prediction Network (HDC-Net), which fuses convolutional and recurrent architectures into a unified framework. HDC-Net employs dilated Convolutional Neural Networks (CNNs) to extract spatial context, together with a novel dual-branch recurrent module combining Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) layers to capture both short-term and long-term motion dynamics. A self-attention mechanism filters out sensor noise and highlights salient features, and a model-agnostic meta-learning (MAML) adaptation scheme enables rapid fine-tuning to new driving scenarios. To address data scarcity in rare crash cases, a Generative Adversarial Network (GAN) synthesizes realistic accident trajectories. Hierarchical attention pooling and convolutional kernel fusion are introduced to optimize inference speed for real-time operation. We evaluate HDC-Net on the recent DeepAccident benchmark with 4-fold cross-validation. HDC-Net achieves 89.3% collision prediction accuracy (CPA), with a 0.42 s time-to-collision error (TCE) and 0.17 m trajectory deviation (TD), while processing each frame in 18.4 ms. It outperforms baseline models and generalizes well to unseen crash types as dataset size grows. These results indicate that integrating dual-branch spatiotemporal learning with noise-aware attention and adaptive meta-learning significantly improves real-time collision forecasting in autonomous systems.

Key words: Collision Detection; Trajectory Tracking; Deep Learning; Meta-Learning; Motion Prediction; Autonomous Vehicles; Sensor Fusion; Computational Optimization