Skip to main content
QUICK REVIEW

[论文解读] The Interstate-24 3D Dataset: a new benchmark for 3D multi-camera vehicle tracking

Derek Gloudemans, Yanbing Wang|arXiv (Cornell University)|Aug 28, 2023
Video Surveillance and Tracking MethodsComputer Science被引用 8
一句话总结

Introduces I24-3D, a 3D multi-camera vehicle tracking dataset recorded along I-24 with 16–17 cameras, 877k 3D vehicle annotations, and baseline benchmarking of multi-camera tracking pipelines.

ABSTRACT

This work presents a novel video dataset recorded from overlapping highway traffic cameras along an urban interstate, enabling multi-camera 3D object tracking in a traffic monitoring context. Data is released from 3 scenes containing video from at least 16 cameras each, totaling 57 minutes in length. 877,000 3D bounding boxes and corresponding object tracklets are fully and accurately annotated for each camera field of view and are combined into a spatially and temporally continuous set of vehicle trajectories for each scene. Lastly, existing algorithms are combined to benchmark a number of 3D multi-camera tracking pipelines on the dataset, with results indicating that the dataset is challenging due to the difficulty of matching objects traveling at high speeds across cameras and heavy object occlusion, potentially for hundreds of frames, during congested traffic. This work aims to enable the development of accurate and automatic vehicle trajectory extraction algorithms, which will play a vital role in understanding impacts of autonomous vehicle technologies on the safety and efficiency of traffic.

研究动机与目标

  • Provide a high-quality, road-traffic 3D multi-camera dataset for vehicle tracking in a real-world highway setting.
  • Enable research on cross-camera 3D tracking, re-identification, and trajectory fusion in congested traffic.
  • Benchmark existing multi-camera tracking pipelines to establish a baseline difficulty and inform future method development.

提出的方法

  • Collect and annotate 3D vehicle bounding boxes across 16–17 overlapping highway cameras over 3 scenes totaling 57 minutes.
  • Annotate in roadway-aligned coordinates with 3D bounding boxes and IDs for re-identification and tracking across cameras.
  • Benchmark multiple existing 3D MOT pipelines by varying detectors, trackers, and cross-camera fusion strategies.
  • Use 3D detections from monocular detectors, multiple tracking/association methods, and two fusion strategies (detection fusion and trajectory fusion) to create cross-camera tracks.
Figure 1: Example annotated (green boxes) frames from each camera field of view for one scene of the I24-3D Dataset. The approximate field of view for each camera is shown on the overhead roadway diagram below (some cameras shown in unique colors as examples). Regions outside of the considered field
Figure 1: Example annotated (green boxes) frames from each camera field of view for one scene of the I24-3D Dataset. The approximate field of view for each camera is shown on the overhead roadway diagram below (some cameras shown in unique colors as examples). Regions outside of the considered field

实验结果

研究问题

  • RQ1Can 3D multi-camera tracking be effectively performed on overhead highway camera networks with precise roadway coordinates?
  • RQ2How do current 3D MOT pipelines perform on densely occluded, high-speed traffic across many cameras?
  • RQ3What is the impact of cross-camera fusion strategies on 3D tracking performance in a traffic monitoring context?

主要发现

  • I24-3D provides 877k 3D vehicle bounding boxes, 720 unique vehicle trajectories, across 16–17 cameras in 3 scenes.
  • Best performing pipeline (Dual3D detector + KIOU tracker + trajectory fusion) achieves HOTA 44.8% and mostly tracked objects 63.8% with GT inputs yielding higher HOTA (59.6% with GT detections, 61.6% with GT tracklets).
  • Even with ground-truth inputs, HOTA remains below 0.6 in many settings, indicating substantial cross-camera tracking difficulty due to localization accuracy and occlusion in dense traffic.
  • Localization errors are non-negligible (average 1.24 ft positional discrepancy across cameras; 0.5 ft average dimension error).
  • Cross-camera rectification and fusion significantly affect performance; trajectory fusion often yields strong results, but no method achieves fine-grained traffic analysis performance (e.g., HOTA > 0.75, MT/ML balanced).
  • The dataset reveals substantial challenges in tracking all vehicles in congested scenes, especially with occlusions and faster-moving objects.
Figure 2: Example single annotation. The annotation is stored in roadway coordinates (left) but can be projected into cameras 5 and 6 on pole 1 (p1c5 and p1c6).
Figure 2: Example single annotation. The annotation is stored in roadway coordinates (left) but can be projected into cameras 5 and 6 on pole 1 (p1c5 and p1c6).

阅读本论文

在 Nubint 阅读器中打开原文,边读边就任意段落提问。

在 Nubint 中阅读本论文

本解读由 AI 根据论文原文生成,引用前请核对原文。