Mapless Yet Accurate: Trajectory Prediction for traffic agents using online HD Map Reconstruction for Autonomous Driving

Accurate trajectory prediction of traffic agents is vital for enabling safer and more reliable autonomous driving. High-definition (HD) and standard-definition (SD) maps play a critical role in this process by providing detailed lane topology and directional cues essential for forecasting the movement of surrounding traffic agents. However, creating HD maps is both expensive and resource-intensive, often relying on complex SLAM systems and specialized sensors, while SD maps, though more readily available, lack the precision needed for accurate navigation in autonomous driving. In this work, we present a novel framework for trajectory prediction that uses online reconstruction of HD maps using images from vehicle-mounted cameras, offering a more scalable and cost-effective solution. Our method achieves notable improvements in prediction accuracy, especially in scenarios lacking access to pre-built maps. Additionally, we propose a new evaluation metric that emphasizes safety by incorporating heuristic weights based on agent relevance and distance into traditional metrics such as ADE, FDE, and Brier-minFDE. This safety-aware evaluation provides a more domain-specific assessment of performance, reinforcing the importance of safety in autonomous driving and establishing a new benchmark for trajectory prediction.