TY - GEN
T1 - Enhance Radar Point Cloud with 2D Diffusion
AU - LI, Yinbao
AU - ZHANG, Yulei
AU - ZHU, Rui
AU - LIU, Chang
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.
PY - 2025
Y1 - 2025
N2 - We propose a novel local diffusion aided single stage detector for radar to tackle the noise and sparsity issues of its point cloud in detection tasks. We utilized space occupancy maps to represent the downsampled point cloud and applied a diffusion module to denoise them. We also introduced instance-aware downsampling strategies and 3D upsampling module to enhance the model’s perception ability in 3D space. Experiments show that LDRadSSD outperformed those SOTA approaches in predicting bounding boxes for road users such as cyclists and pedestrians and figuring out the drivable space in bird’s eye view (BEV) of the scene. In particular, with IoU thresholds of 0.5/0.25/0.25, the average prediction precision (AP) of main type of road users (cars, pedestrians and cyclists) reached at competitive 58.5%, 63.1%, and 82.1%, respectively, while mean IoU of free space was 92.6%. Moreover, the prediction precision of object orientation dramatically raised to averaged 71.1%. We also demonstrate in this paper that LDRadSSD can satisfy real-time requirements in autonomous driving as its running speed reach at 18.2 to 41.7 milliseconds per frame.
AB - We propose a novel local diffusion aided single stage detector for radar to tackle the noise and sparsity issues of its point cloud in detection tasks. We utilized space occupancy maps to represent the downsampled point cloud and applied a diffusion module to denoise them. We also introduced instance-aware downsampling strategies and 3D upsampling module to enhance the model’s perception ability in 3D space. Experiments show that LDRadSSD outperformed those SOTA approaches in predicting bounding boxes for road users such as cyclists and pedestrians and figuring out the drivable space in bird’s eye view (BEV) of the scene. In particular, with IoU thresholds of 0.5/0.25/0.25, the average prediction precision (AP) of main type of road users (cars, pedestrians and cyclists) reached at competitive 58.5%, 63.1%, and 82.1%, respectively, while mean IoU of free space was 92.6%. Moreover, the prediction precision of object orientation dramatically raised to averaged 71.1%. We also demonstrate in this paper that LDRadSSD can satisfy real-time requirements in autonomous driving as its running speed reach at 18.2 to 41.7 milliseconds per frame.
KW - Diffusion model
KW - Object detection
KW - Radar point cloud
UR - https://www.scopus.com/pages/publications/105009978857
U2 - 10.1007/978-981-96-6606-5_2
DO - 10.1007/978-981-96-6606-5_2
M3 - Conference paper (refereed)
AN - SCOPUS:105009978857
SN - 9789819666058
T3 - Lecture Notes in Computer Science
SP - 17
EP - 31
BT - Neural Information Processing : 31st International Conference, ICONIP 2024, Auckland, New Zealand, December 2–6, 2024, Proceedings, Part XI
A2 - MAHMUD, Mufti
A2 - DOBORJEH, Maryam
A2 - WONG, Kevin
A2 - LEUNG, Andrew Chi Sing
A2 - DOBORJEH, Zohreh
A2 - TANVEER, M.
PB - Springer Science and Business Media Deutschland GmbH
T2 - 31st International Conference on Neural Information Processing, ICONIP 2024
Y2 - 2 December 2024 through 6 December 2024
ER -