当前位置: X-MOL 学术Autom. Constr. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Tracking multiple construction workers through deep learning and the gradient based method with re-matching based on multi-object tracking accuracy
Automation in Construction ( IF 9.6 ) Pub Date : 2020-11-01 , DOI: 10.1016/j.autcon.2020.103308
Ohay Angah , Albert Y. Chen

Abstract Multiple construction worker tracking is an active research area critical to the planning of the job site. Challenges in multiple construction worker tracking include miss detection and mismatch due to occlusion and identity switches. To the best knowledge of the authors, the mismatch is not reported in the literature of construction for image based single camera multiple worker tracking. As a result, the mismatch should be taken into account through a representative performance index such as the Multi-Object Tracking Accuracy (MOTA). This work aims to improve the performance of the current multiple worker tracking through an approach composed of three stages: detection, matching and re-matching. In the detection stage, the deep learning detector, Mask R-CNN, is utilized. In the matching stage, we attempt to track workers between consecutive image frames through a gradient based method with feature based comparison. Several cost means and matching methods have been experimented for model selection. Trajectories of tracking objects are derived in this stage. The best cost measurements and matching methods are recommended. Trajectories of tracking objects could be interrupted because of miss detection or mismatch. We call those broken trajectories, without matched detections, orphans. In the re-matching stage, we attempt to recover unmatched detections in the current frame with previous orphans based on extracted features. A competitive MOTA of 56.7% was obtained from the proposed approach over MOTA of 55.9% from the state-of-the-art Detect-And-Track model on a human tracking benchmark dataset. On construction job sites, we have tested the approach with 4 testing videos, resulting in a total MOTA of 81.8%, average MOTA per video of 79.0% and standard deviation of 13.0%, while the maximum and minimum MOTAs are 96.0% and 69.0%, respectively. As a result, the proposed work could potentially provide better multiple worker tracking on the construction job site. Additionally, to have a better representation of the tracking errors, this work suggests to utilize the MOTA for multiple construction worker tracking.

中文翻译:

通过深度学习和基于多目标跟踪精度重新匹配的基于梯度的方法跟踪多个建筑工人

摘要 多个建筑工人跟踪是一个活跃的研究领域,对工作现场的规划至关重要。多个建筑工人跟踪的挑战包括由于遮挡和身份切换导致的漏检和不匹配。据作者所知,基于图像的单相机多工人跟踪的构建文献中没有报告不匹配。因此,应通过多目标跟踪精度 (MOTA) 等代表性性能指标来考虑不匹配。这项工作旨在通过由三个阶段组成的方法来提高当前多工人跟踪的性能:检测、匹配和重新匹配。在检测阶段,使用深度学习检测器 Mask R-CNN。在匹配阶段,我们尝试通过基于梯度的方法和基于特征的比较来跟踪连续图像帧之间的工作人员。已经为模型选择试验了几种成本手段和匹配方法。在此阶段导出跟踪对象的轨迹。推荐最好的成本测量和匹配方法。由于漏检或不匹配,跟踪对象的轨迹可能会中断。我们将那些没有匹配检测的破碎轨迹称为孤儿。在重新匹配阶段,我们尝试根据提取的特征恢复当前帧中与先前孤儿的未匹配检测。56.7% 的竞争 MOTA 是从人类跟踪基准数据集上最先进的 Detect-And-Track 模型的 MOTA 55.9% 的建议方法中获得的。在建筑工地上,我们用 4 个测试视频对该方法进行了测试,结果总 MOTA 为 81.8%,每个视频的平均 MOTA 为 79.0%,标准差为 13.0%,而最大和最小 MOTA 分别为 96.0% 和 69.0%。因此,拟议的工作可能会在施工现场提供更好的多人跟踪。此外,为了更好地表示跟踪错误,这项工作建议利用 MOTA 进行多个建筑工人的跟踪。
更新日期:2020-11-01
down
wechat
bug