ITSC 2025 Paper Abstract

Paper TH-LM-T17.6

Shinoda, Haku (Kanazawa University), Bok, Yunsoo (Kanazawa University), Yoneda, Keisuke (Kanazawa University), Hariya, Keigo (Kanazawa University), Yukiya, Fukuda (Kanazawa University), Suganuma, Naoki (Kanazawa University)

A Simple Pipeline for 3D Object Detection and Motion Prediction without Prior Information through Semantic Understanding

Scheduled for presentation during the Invited Session "S17a-Synthetic-Data-Aided Safety-Critical Scenario Understanding in ITS" (TH-LM-T17), Thursday, November 20, 2025, 12:10−12:30, Southport 2

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 Deep Learning for Scene Understanding and Semantic Segmentation in Autonomous Vehicles, Real-time Object Detection and Tracking for Dynamic Traffic Environments

Abstract

Object detection and motion prediction of surrounding traffic participants are crucial for safe driving of autonomous vehicles. In general, numerous approaches for motion prediction rely on tracking information obtained through object detection and tracking process and improve prediction accuracy by utilizing high-definition (HD) maps. However, dependences on these elements may lead to a reduction in prediction performance in real-world driving scenarios, particularly in map-less environments or under tracking inaccuracies. As an alternative approach, we propose a simple pipeline for 3D object detection and motion prediction only using sequential LiDAR point clouds. The proposed method performs multi-task without prior information such as HD maps or tracking results, by extending PointPillars, a baseline 3D object detection network. Evaluations on our dataset show that detection accuracy of small objects such as cyclists and pedestrians improved by utilizing sequential point clouds. In motion prediction, the proposed method also achieved higher accuracy than the Constant Acceleration model across all the classes. In particular, Self-Attention Blocks and the MapSeg module respectively enhance temporal feature extraction and road structure learning, both contributing to improved prediction accuracy for moving objects.