ITSC 2025 Paper Abstract

Paper FR-LM-T31.6

Aditya, Wisnu (Feng Chia University), Chen, Wen-Hui (National Taipei University of Technology), Lin, Yu-Chen (Feng Chia University), Yang, Ting-Wei (Feng Chia University)

LightMotion: A Lightweight Motion Prediction Model with Spatial Attention and Progressive Learning

Scheduled for presentation during the Regular Session "S31a-AI-Driven Motion Prediction and Safe Control for Autonomous Systems" (FR-LM-T31), Friday, November 21, 2025, 12:10−12:30, Southport 1

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Keywords Advanced Sensor Fusion for Robust Autonomous Vehicle Perception, Deep Learning for Scene Understanding and Semantic Segmentation in Autonomous Vehicles, Real-time Object Detection and Tracking for Dynamic Traffic Environments

Abstract

Achieving an accurate end-to-end motion prediction model with low computational cost remains a challenge. While existing state-of-the-art models deliver strong performance, they rely on computationally expensive architectures. We propose LightMotion, an efficient motion prediction model that maintains high accuracy while significantly reducing computational cost. LightMotion introduces two key innovations: (1) a FLOPefficient ResNet-50 backbone with a post-Feature Pyramid Networks (FPN) spatial attention module that recovers lost resolution in a cost-constrained pipeline and (2) a dual-stage feature recovery mechanism combining spatial attention and progressive learning to refine spatial-temporal cues for more effective trajectory modeling. The attention module enhances critical spatial features at minimal cost, while the progressive mechanism incrementally refines degraded spatial-temporal cues, enabling more effective downstream trajectory modeling. Empirical results show that LightMotion achieves competitive prediction accuracy compared to heavier baselines, operating with up to 67% reduction in FLOPs and only marginal accuracy drops in detection, with comparable results in motion prediction and planning on the Nuscenes dataset benchmark. These results suggest that LightMotion is an efficient solution for motion prediction, balancing low computational cost with high accuracy.