Single-Frame LiDAR Bird's-Eye View with Lightweight Transformer for Self-Driving Vehicle Trajectory Prediction Baseline Research

-- Open-source Lightweight Solution for Mapless Urban Driving

Abstract

Urban advanced driver-assistance systems (ADAS) need to predict the future trajectory of the ego-vehicle at a frequency of 10–20 Hz. Traditional methods rely on high-definition maps, which have pain points such as low freshness, high cost, and narrow coverage. This paper proposes an open-source ego-vehicle trajectory prediction baseline that uses only single-frame LiDAR and requires no high-definition maps or visual assistance. A 200×200×8 bird's-eye view pseudo-image is generated by height-sliced voxelization, and end-to-end trajectory regression is completed by a 6-layer Transformer encoder and uniform Token sampling. After 70% sparse pruning and INT8 quantization, the model size is compressed to 8 MB, and 20 future waypoints are output within 30 ms on Jetson Orin. The ADE on the nuScenes-mini test set reaches 0.54 m, which is comparable to the visual BEV baseline, with a parameter size of <12 M. The code and weights have been open-sourced to facilitate subsequent fusion and temporal expansion.