Robot Service Tasks-oriented Family Environment Cognition Mechanism by Deep Learning*

摘要

In order to improve cognition of service robots, a novel cognition mechanism of family environment based on deep learning is presented in this paper. Facing to an unexplored family environment, the robot can comprehend circumstances efficiently and autonomously with the cognition mechanism rather than the more participation of manual work. The cognition mechanism transforms environment vision information into environment knowledge, guiding robots to perform service tasks. Two designed models play an import role in the cognition mechanism, which are multi-stage feature fusion based on ResNet model (MSFF-R) for family scene recognition and single shot multibox detector increase model (SSD-I) for object detection. The outputs of the MSFF-R model and SSD-I model are semantic descriptions of current surroundings, which are fused to be stored in a designed knowledge base of family environment. The knowledge base can offer the knowledge related to service tasks to guide robot and it can be updated dynamically with the changes of the environment. Finally, MSFF-R and SSD-I model are trained on public datasets and experiments in a practical environment have been done to prove the feasibility of the cognition mechanism.