Autonomous driving in complex urban environments faces significant challenges due to dynamic obstacles, sensor occlusions, and unpredictable interactions among multiple agents. Current obstacle detection systems often struggle with robustness under adverse conditions, while trajectory prediction algorithms frequently prioritize short-term accuracy over long-term temporal reliability. These limitations hinder the safe and efficient deployment of autonomous vehicles in real-world urban scenarios, necessitating a more integrated approach that balances perception accuracy with real-time constraints. This study proposes a unified framework combining multi-sensor fusion for obstacle detection and time-constrained trajectory prediction to address these challenges. The methodology integrates LiDAR and camera data through an optimized neural architecture, enhancing detection robustness against occlusions and sensor noise. Additionally, the trajectory prediction model introduces dynamically adjusted time windows, formulated as a constrained optimization problem using simulated annealing to ensure both spatial accuracy and temporal feasibility. Experimental validation in a CARLA-ROS co-simulation environment demonstrates the system’s effectiveness, achieving an 84.1% recall rate under heavy occlusions and reducing time window violations by 67.3% compared to conventional methods. The framework maintains real-time performance with an average latency of 82.1 ms, making it suitable for urban autonomous driving applications. The research contributes a scalable solution that improves both perception reliability and decision-making safety in dynamic urban environments. By explicitly incorporating temporal constraints into trajectory prediction, the proposed approach enhances the practical deployability of autonomous systems, paving the way for more adaptive and robust navigation technologies in complex traffic scenarios. Future extensions could explore adaptive parameter tuning and cooperative prediction in connected vehicle environments.
Jiawang Tan (Thu,) studied this question.