Pedestrian Trajectory Prediction Based On Part Affinity Fileds And Long Short Term Memory Network

As a vulnerable road user in a mixed traffic environment,the prediction of pedestrians’ trajectory is very important for autonomous vehicles or intelligent auxiliary traffic systems.Pedestrian trajectory prediction is a task of predicting the future trajectory of a pedestrian based on its past trajectory.Autonomous vehicles deployed on urban roads should ensure the safety of road users and be able to drive effectively without obstructing traffic,which requires autonomous vehicles to extrapolate the future movements of nearby traffic participants.Because of the great subjectivity and randomness in the pedestrian movement,the prediction of its trajectory has become a challenging research hotspot.Based on the pose estimation of human body using part affinity fields,this paper identifies the pedestrian’s movement intention,and finally uses the Long Short Term Memory Network integrating attention mechanism to predict the pedestrian’s movement trajectory in a short time.The main research contents of this paper include:1)Research on human pose estimation using Part Affinity Fields.Deep convolutional network was used to process the human body image to obtain the confidence map（to obtain the position information of each joint）containing 18 skeleton key points of the human head,shoulder,wrist,ankle,etc.Using the method of part affinity fields（obtaining limb position and direction information）to analyze the situation of multiple people in the image,the affinity fields of 17 groups of connected skeleton key points,including right shoulder to right elbow,right elbow to right wrist,etc,was obtained.Based on the affinity domain information between the confidence map of skeleton key points and the connected key points,the Hungarian algorithm in the maximum weight binary graph matching problem is used to correlate the skeleton key points belonging to the same person,and these key points are connected into a complete human pose skeleton.2)Research on pedestrian movement intention recognition based on human pose estimation.This paper classifies pedestrians intention in road traffic environment into four categories: Starting,Walking,Stopping and Standing.According to the distribution features of pedestrian skeleton key points and its relation with the connotation of pedestrian movement pose,build the pedestrian pose feature vector and the first-order Hidden Markov Model is used to realize（Hidden Markov Model,HMM）pedestrian movement intent type conversion between modeling,and uses the viterbi algorithm to find the most likely sequence of hidden states,identification of pedestrian movement intentions.3)Research on pedestrian trajectory prediction based on Long Short Term Memory Network.According to the pedestrian’s historical trajectory sequence,motion intention and pedestrian scale features,the Long Short Term Memory Network model in deep learning is used to predict the pedestrian’s motion trajectory.In order to depict the interaction between pedestrians,this paper also uses the attention mechanism to select the location information of other individuals around the target pedestrian and input it into the Long Short Term Memory Network,so as to make a more consistent social norms and more reasonable prediction of the pedestrian’s movement trajectory.Finally,two evaluation indexes,Final Displacement Error（FDE）and Average Displacement Error（ADE）,were used to compare and analyze the proposed method with other three different trajectory prediction methods.The experimental results show that the pedestrian trajectory prediction algorithm based on Part Affinity Fields and Long Short Term Memory Network has the highest accuracy,good performance and certain applicability in both indexes.