Roadside Pedestrian Motion Prediction Approach

Accidents between vehicle and pedestrian account for a large partition of whole traffic accidents,and pedestrian motion prediction is a major concern for intelligent vehicle in the near future,which is significant for reducing accidents and improving traffic safety.However,current research is poor in some aspects,for example,some failed to use pedestrian pose and traffic condition for prediction;relative research are scattered and divided into too many small fields,thus no integrated system is proposed to connect perception and decision.To address aforementioned issues,this thesis takes pedestrian pose and traffic environment into consideration and proposes a method for pedestrian motion prediction for pedestrian crossing scenarios without traffic light and intersection.Firstly,an algorithm for pedestrian behavior estimation based on Bayesian estimation and an algorithm for pedestrian crossing intention estimation based on Dynamic Bayesian Network are proposed successively.The former algorithm defines ‘stand',‘walk' and ‘run' as three typical behaviors,then build a prior probability model using Long Short Term Memory model and a maximum likelihood probability model using distribution of pedestrian velocity to estimate the posterior probability of pedestrian behavior.The latter algorithm designs a Dynamic Bayesian Network for the inference of pedestrian crossing intention,which both takes the pedestrian goal,scenario criticality etc.from pedestrian view as latent variables and pedestrian body orientation,relative longitudinal distance etc.from vehicle view as observable variables.Secondly,based on the probability estimation of pedestrian behavior and crossing intention,an algorithm for pedestrian trajectory prediction imitating particle filter is proposed.For the development of the algorithm,a grid map is defined by analyzing the key elements in the traffic environment;by ananlyzing the characteristics of pedestrian movement and the shortcums of traditional motion models,a pedestrian motion model is proposed;particles are used to simulate pedestrian movements by considering pedestrian behavior,and the importance sampling is done considering the environment and pedestrian crossing intention,thus predicted trajectories are given with distribution of particles.Thirdly,an integrated hardware and software platform for pedestrian and environment data collection is set up based on essential sensors and Robot Operationg System.With the help of some open source algorithms and labelling tools,a method for data processing is proposed to provide pedestrian key points,pedestrian body orientation,pedestrian position,road structure and ego vehicle motion information as the base of pedestrian motion prediction algorithm.Finally,eight typicle scenarios are defined by road type,vehicle velicity etc.and data collection is done with test vehicle.These data are used to verify the effectiveness of preposed algorithms.The results show the accuracy of the prior model for behavior estimation is more than 0.97,the maximum likelihood model is more than 90% and the posterior probability model is also nore than 0.98;the algorithm for pedestrian crossing intention estimation is able to estimate pedestrian's intention to cross at a relative distance of about 20 m and 0.2~0.5s earlier before pedestrian truly crosses;the algorithm for trajectory prediction also gives a distribution of pedestrian position in future 3 seconds,which matches the pedestrian behavior and intentions well.