Trajectory Prediction for Collision Avoidance of Vulnerable Road Users in CCAM Scenarios Using Stacking Regression and Deep Learning

Road accidents remain a critical concern within the Cooperative, Connected, and Automated Mobility (CCAM) ecosystem, particularly for Vulnerable Road Users (VRUs) such as cyclists, moped riders, and motorcyclists. These users typically lack access to advanced onboard sensors, yet often carry electronic devices capable of emitting valuable geospatial data. This chapter proposes a hybrid trajectory prediction framework based on stacking regression and deep learning models to enhance collision avoidance systems within CCAM operational domains such as urban driving and protected zones. The methodology leverages lightweight models optimized for real‑time inference, ensuring low latency and minimal energy consumption, which are key requirements in CCAM applications. A large‑scale SUMO–TraCI co‑simulation is used to generate realistic trajectories for thousands of road users, including VRUs, under heterogeneous urban conditions with signalized intersections and risky behaviours. From these simulations, time‑windowed geospatial sequences of latitude, longitude, and heading are constructed to train and compare multiple regression and deep learning architectures. Performance is evaluated in terms of prediction error over a 5‑second horizon, execution time, and estimated CO₂ emissions. The best‑performing stacking regressor and LSTM‑based model are then integrated into a geometric collision detection module based on polygon intersection and a Monte Carlo algorithm for probabilistic risk estimation. Experimental results show improvements in execution time and computational efficiency compared to traditional baselines, while maintaining competitive prediction accuracy. The proposed framework supports earlier and more reliable collision warnings for VRUs in CCAM environments and promotes a sustainable approach that balances safety performance with computational and environmental costs.