Research On Behavior Recognition For Home Service Robot

With the continuous breakthrough and progress in communication,artificial intelligence,machine vision and other fields,the related research on robot technology continues to develop to a deeper and higher level.As an intelligent robot oriented to the family working environment,household service robot has become an inevitable development trend in the future.How to make the robot recognize the human posture and movement effectively and help it to better actively"perceive" and "recognize" human behavior has been an important issue in the academic field.Different from the morphological detection in a single picture,the behavior recognition task mainly explores how to perceive the action changes of an object or even multiple objects in a continuous video stream,and then increases from the summary of continuous actions to the judgment of behaviors.With the rapid development in recent years,deep learning has achieved good results in many visual fields,but there are still many problems in deploying deep learningbased behavior recognition algorithms to mobile robot platforms.Because of its method nature,deep learning method requires high computing power on hardware,especially in the processing of image tasks,which requires a large amount of GPU graphics.Therefore,how to reduce the computational load and transplant the algorithm to the home service robot platform on the premise of ensuring the speed and accuracy of the algorithm is still a big challenge.This paper focuses on the behavior recognition of household service robots,proposes a behavior recognition algorithm for household service robots,and studies the related technologies in the behavior recognition algorithm.The main research contents of this paper are as follows:(1)the proposal and establishment of special data sets for behavior understanding of household service robots.By analyzing and summarizing the common behaviors of household service robot in daily environment,the common behaviors are put forward.The video was collected according to the behavior category and the data set was established.The data set format of the current mainstream open data set was used to complete the production and annotation.(2)proposed behavior recognition algorithm based on compressed video.By comparing the existing behavior recognition algorithms and combining the design requirements and working environment of household service robots,a behavior recognition algorithm for household service robots is proposed.In addition,in order to apply to the mobile robot platform,the requirements of hardware performance of the algorithm are reduced as much as possible,and the running speed of the algorithm is accelerated.The algorithm is tested and evaluated by experiments on the open data set and the self-built dataset.(3)the behavior recognition algorithm model is compressed and accelerated.On this basis,a model compression algorithm based on dynamic filter pruning is proposed,which accelerates the inference speed of the model with a small accuracy loss,making it suitable for the domestic service robot platform with insufficient computational force and high real-time requirement.