Overcoming Challenges in Motion Forecasting for Autonomous Vehicles: A Comprehensive Analysis
Year: 2025
Type: article
Abstract: Motion forecasting is crucial for Autonomous Vehicles (AVs) to navigate dynamically safely by predicting the motion of surrounding agents. Significant challenges include fusing heterogeneous data (e.g., HD maps, historical trajectories, real-time sensor readings), modeling dynamic interactions between agents and between agents and dynamic/static worlds, and handling multimodal behavior forecasting from numerous driving scenarios. Additional challenges include ensuring model robustness with partial sensor data and increasing data interpretability of data driven solutions. The new proposals offer hybrid systems combining scenario-based and perception based methodologies. For example, Real Motion employs temporal context and spatial relations to enhance prediction accuracy through the fusion of historical driving behavior and real-time scene analysis. Advanced techniques like attention mechanisms and graph neural networks increase interaction modeling, and hybrid data-driven/model-based techniques find a compromise between common-scenario performance and rare-event management. This research has the objective of improving the ability of autonomous vehicle motion prediction models to generalize in order to reduce the performance loss in unseen or adverse environments and also elucidate future research direction.
Authors: Dr. Ravi Khurana, Arti, Mandip Kaur
Institution: Kanya Maha Vidyalaya
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4410814207