Cloud-Based Knowledge Sharing in Cooperative Robot Tracking of Multiple Targets with Deep Neural Network

摘要

Cooperative robot tracking of multiple targets plays an important role in many realistic robot applications. In order to minimize the time during which any target is not tracked, target trading among robots at runtime is a common phenomenon. After a period of successful tracking, the robot can gain a lot of knowledge about the target details, for example, the appearance changes caused by motion and illumination. However, the accumulated knowledge is dropped simply in existing research while robots trading targets, which makes each robot has to learn the knowledge of target details from scratch. The absence of knowledge sharing heavily influences the tracking accuracy in practice. In this paper, we propose a novel approach named Cloudroid Tracking which enables knowledge sharing through the support of the back-end cloud infrastructure. Our approach adopts the deep neural network (DNN) and its online tuning mechanisms to enable the knowledge accumulation. The dynamic connection of multiple DNNs on the cloud infrastructure and multiple robots is enabled. No matter how the target changes, the robot can connect to the corresponding neural network which is responsible for a specific target. The experimental results on both open dataset and real robots show that our approach can promote the accuracy for robot tracking significantly.