Pedestrian Tracking in Public Passageway by Single 3D Depth Sensor
Pedestrian Tracking in Public Passageway by Single 3D Depth Sensor
复制标题
DOI：
10.1109/percomworkshops53856.2022.9767224
复制
发表时间：
2022-03
期刊：
影响因子：
--
通讯作者：
Riki Ukyo;Tatsuya Amano;Akihito Hiromori;Hirozumi Yamaguchi
中科院分区：
文献类型：
--
作者：
Riki Ukyo;Tatsuya Amano;Akihito Hiromori;Hirozumi Yamaguchi
我们提出了一种基于单个3D深度传感器捕获的3D点云数据的行人跟踪方法，该方法可以在公共通道中跟踪各种行人，例如携带行李和婴儿车的行人以及彼此靠近的家庭团体。由于我们假设传感器附着在墙壁上，很容易部署在通道中，因此在传感器附近行走的行人经常会遮挡后面的其他人。这在三维点云和基于卡尔曼滤波的跟踪中会导致严重的行人分割错误。为了解决这一问题，我们引入了一种新的基于kf的多目标跟踪中缺失部分的空间补充技术。我们使用3D点云数据来评估我们的方法，这些数据捕获了现有商业设施（小型购物中心）入口处的行人，以及我们实验室空间中收集的行人。因此，多目标跟踪精度指数（MOTA）为0.914，遮挡严重且频繁。
We propose an approach to pedestrian tracking in a public passageway with various pedestrians, such as those carrying luggage and baby strollers and family groups close to each other, based on 3D point cloud data captured by a single 3D depth sensor. Since we assume a wall-attached sensor, which is easy to deploy in passageways, pedestrians walking nearby the sensor frequently occlude the others behind. This causes a severe error in pedestrian segmentation in the 3D point cloud and Kalman-filter-based tracking. We introduce a new technique to spatially complement the missing part of segments in KF-based multi-object tracking to cope with this issue. We have evaluated our method using the 3D point cloud data capturing pedestrians at the entrance of an existing commercial facility (shopping small), as well as the one collected in our laboratory space. As a result, the tracking accuracy index (MOTA) for multiple objects is 0.914, with severe and frequent occlusions.