Pedestrian Behaviour Understanding And Robot Behaviour Evaluation For Tri-Co Robot

With an aging population and an increasing labor shortage in services,medical care,education and other industries,the demand for service robots is on the rise.There is a huge market potential and development space for service robots.However,the existing service robots can not satisfy people’s expectations,and Tri-Co Robot(CoexistingCooperative-Cognitive Robot)has become the development direction of next-generation service robots.For service robots working in crowds,improving the navigation ability in a human-robot coexistence environment is the basis of human-robot interaction,human-robot collaboration and other high-level tasks.Compared with traditional mobile robots,the emergence of humans increases the difficulty of deploying service robots in real-world environments.How to represent pedestrians with social attributes during robot navigation tasks is the first problem that needs to be solved to achieve human-robot coexistence.Specifically,humans are different from obstacles in the environment who have subjective feelings and social attributes.Secondly,forecasting the trajectory of pedestrians is an indispensable part of robots working in dynamic scenes,which can alleviate the problems,such as "turning" and "freezing robot." While solving the above problems,considering the human comfort space and predicted trajectory in the motion planning algorithm to generate a trajectory that considers the comfort of pedestrians is the crucial point of human-aware navigation.Lastly,the evaluation of robot behaviour is essential for enhancing robot navigation ability and continuously optimising robot behaviour.Because of these problems,this paper explores the field of human comfort space modelling,pedestrian trajectory prediction,human-aware robot navigation,and robot behaviour evaluation.The main contributions are listed as follows:1)Modeling of human dynamic comfort space based on asymmetric gaussian function: In view of the previous pedestrian comfort space model,it has a fixed shape and can satisfy the comfort needs of various pedestrians.This paper proposes a novel dynamic personal comfort space model based on asymmetric Gaussian function.First,human movement is used to calculate the shape of different comfort spaces.Then a scalable fuzzy reasoning framework is proposed to define the size of personalized comfort space,which considers differences in individual social attributes such as gender and age.When a single human’s comfort space has been modelled,this paper proposes the concept of a common concern area and detects it through the minimum covering circle algorithm to construct the comfort space for the groups.Quantitative and qualitative assessments indicate the effectiveness of the proposed dynamic comfort space.2)Pedestrian trajectory prediction using group constrained hierarchical graph attention networks: Most of the existing interaction models in trajectory prediction only focus on the interaction between pairs of pedestrians and do not consider the group movement of pedestrians and the interaction between groups.Subsequently,this paper proposes a group-constrained hierarchical graph attention network based on generative adversarial networks for pedestrian trajectory prediction.Based on our research accumulation in group detection and group activity recognition,group constraints are introduced on the basis of traditional paired pedestrian interaction modelling,and the interaction between pedestrians is considered from the individual and group levels.The interaction between pedestrians in the scene is split into three parts: inter-group,intragroup,and out-group interaction,and then the hierarchical graph attention network is used to model these three interactions,respectively.At the same time,this paper considers the interaction between pedestrians as a disturbance in the process of pedestrian movement.Through experiments on public datasets,the proposed method significantly improves prediction accuracy and generates future trajectories that satisfy social rules.3)Group aware spatial-temporal transformer for human trajectory prediction:Based on Chapter 3,this paper proposes a group-aware spatial temporal transformer trajectory prediction framework.Aiming at the problem that the model proposed in Chapter 3 requires group detection in advance and does not fully extract the spatial-temporal dependency between temporal and spatial distribution information.In this paper,because of the advanced attention mechanism of the transformer network,the temporal transformer is used to extract the temporal motion features of pedestrians,and the spatial transformer is used to extract the spatial distribution characteristics of pedestrians.Spatial temporal dependence is subsequently captured through the cross-attention mechanism.At the same time,by sharing the motion encoder with the trajectory prediction task,the group detection task is converted into the regression of the adjacency matrix.Then the individual representation is updated with the calculated adjacency matrix.In this way,group information is incorporated into individual representations,and the group constraints are considered in the trajectory prediction process.Lastly,quantitative experimental results indicate that the proposed methods have a 19.4% improvement in ADE compared to the latest baseline method and successfully forecast the group coherent movement of pedestrians.4)Social aware navigation and robot behavior evaluation in the populated environment: This paper proposes an interaction-aware robot navigation framework based on the above research.The framework designs a temporal social cost map to represent pedestrians based on the layered cost map mechanism.Specifically,the comfort space model proposed in Chapter 1 is used to represent pedestrians in the scene and then normalized to the social cost layer.At the same time,according to the historical observation,the trajectory prediction module obtains the T time step future trajectory and then constructs the time-series social cost map for the future T time step.The timeseries social cost map in time dependence A* algorithm can be used as a cost item for evaluating motion primitives.In selecting motion primitives,trajectories that satisfy pedestrian comfort and have a low collision probability are selected.In this way,the robot can understand human interaction when navigating in crowds.In addition,it is not easy to quantitatively evaluate the comfort level of the robot’s movement.This chapter uses manually annotated public datasets,takes people and robots as nodes in the graph,and then uses graph neural networks to aggregate the information of each node.The characteristics of the entire scene are obtained to evaluate the discomfort caused by the presence of the robot in the current scene.Since the dataset is manually annotated,there are people involved in the evaluation process,so the annotation of the dataset includes human cognition of comfort.Compared with the traditional robot behaviour evaluation method without considering human subjectivity,the evaluation network trained according to the manually labelled dataset includes human subjectivity.It is closer to the real feelings of humans.This paper first completes the personalized representation of pedestrian social attributes through human comfort space modelling.Subsequently,by exploiting the characteristics of pedestrian movement,group constraints are introduced into pedestrian interaction modelling for trajectory prediction,which significantly improves prediction accuracy.The comfort space model and pedestrian trajectory prediction module are applied to the motion planning algorithm to generate socially accepted trajectories.In this way,the robot can understand human interaction when navigating in crowds.Lastly,the robot behaviour evaluation network is trained by manually annotated data sets and integrates human cognition during the evaluation process.