PERIAPT: Joint Person Detection, Re-Identification and Pose Tracking in Video

对于机器人、自动驾驶或智能工厂等许多应用来说，机器必须意识到附近的人类。这需要知道当前人类的位置和姿势，并持续跟踪他们。虽然人的检测、人的再识别和姿态跟踪是高度相关的，但这三个任务之前都是独立研究的。例如，被检测人员边界框的对齐质量会影响再识别的准确性。如果一个人被遮挡了很长时间或离开现场后再次出现，则需要重新识别。此外，人体检测器很难检测到部分闭塞的人，而即使人体的其他部分闭塞，也可以检测到可见的关节。因此，我们将研究这三个任务如何相互协助，我们的目标是通过在一个联合框架中一起解决这三个任务来提高这三个任务的准确性。

For many applications like robotics, autonomous driving, or smart factories, machines have to be aware of humans in their vicinity. This requires to know the location and pose of the present humans and track them continuously. Although person detection, person reidentification and pose tracking are highly correlated, the three tasks have been previously studied independently. For instance, the alignment quality of the bounding boxes of detected persons affects the re-identification accuracy. If a person is occluded for a long time or re-appears after leaving the scene, it needs to be re-identified. Furthermore, person detectors struggle to detect partially occluded persons whereas visible joints can be detected even if other parts of the human body are occluded. We will therefore investigate how the three tasks can assist each other and we aim to improve the accuracy for all three tasks by solving them together in a joint framework.