Research Of Key Technology In Person Following Robot Based On Deep Learning

Traditional person following robot systems often require well-designed features and controllers to follow an assigned person,and often incorporate multiple sensors to improve robustness,resulting in high design cost and complex multi-sensor data fusion.What’s more,most traditional robot systems fail to deal with target disappearance well and are prone to lose the target.In this paper,Convolutional Neural Network(CNN)and Deep Reinforcement Learning(DRL)in Deep Learning(DL)are introduced into the person following robot system based on monocular vision.The former is responsible for the estimation of the distance between robot and target person,and the latter is responsible for the direction control of the robot.This paper aims to reduce the hardware and design cost of the robot system,improve the system robustness,and reduce the occurrence of robot losing the target person,which is more than necessary for the person following robot systems.To accelerate the convergence of the DRL-based direction control model,we utilize the framework of "supervised learning+model transfer+reinforcement learning(RL)".Firstly,train a CNN-based direction control model with high performance,then transfer the knowledge from the CNN model to a DRL-based direction control model to enable the DRL model to obtain the same direction control ability as the CNN model quickly,and the control ability of the DRL-based model can be enhanced by RL.Experimental results show that the introduction of the RL further improves the model performance,which can be inferred from the model’s correction of target confusion and the fact that robot follows the person farther.Besides,a little data and several times of environmental interaction are enough for the DRL-based model to fully adapt to a new environment.For the training of CNN-based direction control model,this paper proposes an automatic method of dataset construction.After collecting 1000 to 2000 images,a large-scale simulated data set(about 70000 data)can be built automatically with computer program,which significantly reduces the data collection and labelling cost.Experiments show that the simulated data set constructed by this method can reflect the variability and complexity of the real environment well,indicating the high practicability of this method.When it comes to the distance control between the target person and robot,this paper designs a distance control algorithm based on CNN,which integrates the estimation of target position and depth prediction of the scene in traditional algorithm,avoids the manual feature design process,and reduces the requirements on system and environmental conditions,resulting in higher flexibility.Experiments demonstrate that the proposed algorithm can make a reasonable estimation of distance with only monocular images provided,so that robot can adjust speed to keep a safe distance from the target person.The experimental results show that the proposed DL-based person following robot system can adapt well to the environment with such as illumination changes,object interference,target occlusion,pedestrian interference.Furthermore,this system can effectively cope with the situation of target disappearance,which has not been tackled well in most traditional person following robot systems.