Feature-based Knowledge Distillation in 3D Semantic Occupancy Prediction for LiDAR-less Autonomous Driving

Nariki Tanaka･Toshihito Ikenishi (Mitsubishi Electric)

LiDAR-less autonomous driving systems that rely solely on cameras offer several advantages, including lower sensor and computational costs. 3D semantic occupancy prediction is well-suited for these systems because it enables fine-grained scene understanding. However, LiDAR-based models typically achieve higher accuracy than camera-based models. Therefore, we aim to leverage point clouds during training only to improve camera-based models. To this end, we explore cross-modal, feature-based knowledge distillation, a technique widely used in other tasks.

Articulated Vehicle Detection via Learning Inter-Object Relationships using GNN in LiDAR Point Clouds

Riku Kagohashi･Keigo Hariya･Keisuke Yoneda･Yukiya Hukuda･Naoki Suganuma (Kanazawa University)

Detecting articulated vehicles in LiDAR point clouds suffers from low accuracy due to geometric complexity and information loss caused by self-occlusion. This study proposes an object detection model that learns inter-object relationships using Graph Neural Networks (GNN). By jointly considering spatial arrangement and contextual information in addition to individual geometric features, we aim to improve the detection accuracy of articulated vehicles.

Self-Supervised Pre-Training for Generalizable Feature Extraction from 4D Imaging Radar Point Clouds

Takumi Takai･Keisuke Yoneda･Keigo Hariya･Haku Shinoda･Yukiya Hukuda･Naoki Suganuma (Kanazawa University)

4D imaging radars excel in environmental robustness and long-range detection. However, their point clouds contain abundant noise and are sparse, making challenges for tasks such as object detection and self-localization. In this study, we build a self-supervised pre-training model based on a masked reconstruction task to extract generalizable features from radar point clouds, and evaluate its effectiveness in object detection.