Intention Recognition In Human Robot Collaboration

The most paramount issues in process of human robot collaboration are efficiency and security.Recognizing the intention of human worker during collaboration can be a good method to improve the working speed and save time for collaboration,which can also protect human beings from potential danger when cooperating with robot.There are lots of schemes to finish this job,like intention recognition based on emotion recognition,sound track or texting message,body pose or hand gesture,and trajectory prediction.Among them,the recognition method based on the motion trajectory prediction can not only provide the position information of the target object selected by the robot system more intuitively,but also provide more accurate obstacle avoidance information for the collision detection.However,traditional prediction models rely too much on built datasets and are not able to cope with motion trajectories in more complex and diverse collaborative task environments.In response to this problem,this thesis proposes a new solution.First of all,this thesis designs a human robot collaboration system based on the task of HRC.By analyzing the pose of robot and dynamic function,this thesis build up a platform to manipulate the robot,combined with ROS system.Moreover,this thesis proposes a human interactive model based on depth camera.This model comprises of cylinders and spheres representing the human joints.The occupancy volume is calculated and more computation efficiency than other common methods.Meanwhile,this thesis introduces the process of HRC and build up a HRC motion dataset for training models.Then,this thesis studies the current popular neural network structure and method.By comparing neural network models based on different connection methods,it is expected to obtain a prediction model structure suitable for application in a HRC scenario.Most importantly,under the constraints of network structure complexity,this thesis analyzes the mechanism of residual network and proposes a refundable learning method suitable for trajectory prediction.This new connection function can increase the parameter iteration speed of the model and improve the accuracy of the final prediction without increasing the number of connection layers.Moreover,this thesis also proposes two loss functions for estimating the smoothness of the trajectory,which can make the model output a smoother trajectory,making the output data more realistic.In the experiment,this thesis not only tested the proposed method on the dataset based on the new collaborative task,but also verified it in the public data set Human3.6M.As the results show,the method has sufficient robustness and generalization.Finally,in view of the requirement of the recognition task for recognition speed,this thesis proposes a method based on long short-term learning.After analyzing the shortcomings and advantages of long-term prediction and short-term prediction respectively,by combining the two prediction methods,the new scheme adopted in this thesis can double the output trajectory duration without increasing the input trajectory.This avoid the rigid requirement raised by common end-to-end prediction model,which is the number of sampling points for the input and output are supposed to be equal or more than output.Experiments show that although the long short-term learning are not as good as the short-term prediction in the fitting degree,the prediction trend of the trajectory is the same as the ground truth,which can provide the intention of human more quickly and work more efficiently.