BEV-IMTP: Lightweight Multi-Task Learning for Joint Bird's-Eye-View Instance Mapping and Trajectory Prediction

Trajectory prediction is a critical component of autonomous driving systems, enabling vehicles to anticipate the motion of surrounding agents and make safe, informed navigation decisions. This paper presents BEV-IMTP, a lightweight instance mapping and trajectory prediction network designed for Bird’s-Eye-View representations. The framework uses customized core layers and a minimal-parameter backbone to reduce computational overhead for latency-sensitive autonomous driving applications. BEV-IMTP improves semantic map mIoU by 4.6% and instance motion mIoU by 1.8% compared with state-of-the-art methods. On nuScenes, it achieves 62.1% overall semantic map mIoU, with 63.4% for dividers, 58.3% for pedestrian crossings, and 64.6% for boundary regions. The model uses 34.5 million parameters and 88 GFLOPs and reaches 3.1 FPS on a single Tesla V100 GPU. Evaluation on nuScenes and Lyft demonstrates robust generalization across camera configurations.

Discover Artificial Intelligence, 6(1). Springer Nature