Laser & Optoelectronics Progress, Volume. 62, Issue 8, 0815009(2025)
Three-Dimensional Object-Tracking Method Based on Joint of Multidimensional Features of Point Clouds
Figures & Tables(13)
Fig. 7. Effect of different sizes of hyperparameters on AMOTA. (a) Object embedding module; (b) feature matching module; (c) motion pridection module; (d) multi-dimensional feature affinity
Fig. 8. Tracking effect on nuScenes dataset. (a)(d) Proposed algorithm in BEV; (b)(e) CenterPoint in camera view; (c)(f) proposed algorithm in camera view
Fig. 9. Aerial view of tracking effect of validation set. (a) Proposed algorithm; (b) comparison algorithm 1; (c) comparison algorithm 2; (d) comparison algorithm 3; (e) comparison algorithm 4; (f) comparison algorithm 5
Table 1. Experimental software and hardware environment
View tableTable 1. Experimental software and hardware environment
Hardware Specification Software Version CPU Intel Xeon Silver 4214 System Ubuntu 20.04 LTS GPU NVIDIA Corporation GP102GL ［Tesla P40］ GPU×2 Python 3.7.0 PyTorch 1.11.0 Memory 251.5 GiB CUDA 11.3 Hard disk 4.2 TB CUDNN 8.2.1
Table 2. Comparison of experimental results
View tableTable 2. Comparison of experimental results
Algorithm Data type AMOTA /% AMOTP /% MOTA /% IDS AB3DMOT［ 2 ］Lidar 45.4 15.1 15.4 927 Chiu et al.［ 7 ］Lidar 55.0 61.7 45.9 776 Eagermot［ 24 ］Lidar+Camera 66.7 55.0 56.8 1165 Minkowski Tracker［ 25 ］Lidar 67.8 54.0 57.8 325 OGR3MOT［ 26 ］Lidar 65.6 62.7 55.4 288 SimpleTrack［ 9 ］Lidar 66.8 56.8 55.7 609 CenterPoint［ 8 ］Lidar 63.8 55.5 53.7 760 Proposed Lidar 68.2 57.4 55.6 438
Table 3. Different categories track performance
View tableTable 3. Different categories track performance
Algorithm Car Truck Bus Trailer Pedestrian Motorcycle Bicycle AB3DMOT［ 2 ］27.8 1.3 40.8 13.6 14.1 8.1 7.2 SimpleTrack［ 9 ］82.3 58.7 71.5 67.3 79.6 67.4 40.7 OGR3MOT［ 26 ］81.6 59.0 71.1 67.1 78.7 64.0 38.0 Eagermot［ 24 ］58.3 59.7 74.0 63.6 74.4 62.5 58.3 CenterPoint［ 8 ］82.9 59.9 71.1 65.1 76.7 59.1 32.1 Proposed 84.0 53.4 87.8 55.6 78.9 53.6 57.1
Table 4. Ablation experiments
View tableTable 4. Ablation experiments
Method Module AMOTA /% AMOTP /% MOTA /% IDS Embedded and matching network Motions and prediction module Baseline 63.8 55.5 53.7 760 Experiments 1 √ 66.4 47.2 52.4 653 Experiments 2 √ 65.3 54.3 49.7 547 Proposed √ √ 68.2 57.4 55.6 438
Tools
Get Citation
Copy Citation Text
Feng Tian, Sirui Zhang, Fang Liu, Zhuohan Han, Mengyang Zhang, Yizhou Lu, Guibao Ma, Huan Chang, Ling Zhao, Yuxiang Han. Three-Dimensional Object-Tracking Method Based on Joint of Multidimensional Features of Point Clouds[J]. Laser & Optoelectronics Progress, 2025, 62(8): 0815009
Paper Information
Category: Machine Vision
Received: Aug. 19, 2024
Accepted: Oct. 8, 2024
Published Online: Apr. 3, 2025
The Author Email: Fang Liu (Ifliufang1983@126.com)
CSTR:32186.14.LOP241863
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20