A Unified Framework for Pedestrian Trajectory Prediction and Social-Friendly Navigation

In recent years, stable robot navigation systems need to meet the requirements of comfort and sociality, such as maintaining an appropriate distance from pedestrians, avoiding crossing crowds, and so on. However, the traditional robot navigation frameworks treat the surrounding pedestrians or objects as obstacles and fail to solve the navigation problems in the context of human-robot interaction. Therefore, we propose a unified framework for human-aware and social-friendly navigation, which includes three modules: 1) pedestrian modeling, 2) trajectory prediction, and 3) path planning. In this work, we detect and model pedestrians with asymmetric Gaussian function, while introducing motion-consistent feature to identify movement group. For pedestrian trajectory prediction, we propose an efficient and accurate generative adversarial network model, combining social feature attention mechanism, and variable intention filter. For path planning, a “plan-prediction-execution” cycle mode is applied to improve the performance of mobile robots in dynamic environments. The experimental results show that compared with the traditional path planning, our social-friendly navigation framework has higher navigation efficiency and meets the comfort and sociality of social navigation.

I. Introduction

Research on social navigation shows that in order to improve the acceptability of robot navigation in a human-centered environment, robots need to move in an efficient and socially compliant manner that minimizes interference with pedestrians, rather than looking for the shortest collision-free path [1], [2]. However, traditional robot navigation methods treat pedestrians and objects as obstacles, which may cause freezing robot problems [3], and pedestrians may perceive these behaviors as rude or dangerous [4]. In order to enable mobile robots to interact with humans in a natural and friendly way, research projects on social service robots such as SPENCER [5], TERESA [6], and STRANDS [7] have been launched, and some human-aware robot navigation frameworks have been proposed. According to the requirements of pedestrian comfort [8] and sociality [9], the prerequisite for mobile robots to navigate in a socially-compliant manner is to effectively predict pedestrian trajectory. The current trajectory prediction methods mostly focus on modeling the movement of a single pedestrian to predict the future trajectory [10], [11]. However, these trajectory prediction models only consider the spatial position relationship between pedestrians when extracting interaction information, ignoring the influence of pedestrian motion direction, speed on the future trajectory of the target pedestrian. Second, information obtained from previously recorded trajectory data is essential for inferring key characteristics of long-term pedestrian movement, such as pedestrian intention [12], [13]. In addition, Moussaïd et al. [14] showed that there may be social interactions between individuals in a group, which makes it inappropriate for robots to plan a path through the group. Therefore, effective group detection and modeling of pedestrians is required for social-friendly navigation.