Pedestrian Trajectory Prediction Of Variable Length Based On Autoencoder

With the increasing scale of big data,the demand for data analysis is becoming increasingly urgent.Data analysis aims to extract valuable information and knowledge from massive amounts of data.Currently,data analysis for pedestrian activities has important implications for information mining and related research applications in many fields.For example,in the field of autonomous driving,predicting pedestrians’ future behavior can reduce potential risks; in machine navigation,pedestrian future trajectory data can be used to determine whether there is congestion ahead,thus enabling better route selection; in video surveillance and intelligent security systems,pedestrian trajectory information can be used to track people flow and make better plans for fire safety and related matters.However,predicting the future activity trajectory of pedestrians is not an easy task,and there are still several issues in the related work.Firstly,to predict where pedestrians are heading in the future,previous researchers often used only fixed-length trajectory information as input for the entire algorithm model when building relevant models.This approach results in the omission of trajectory data for pedestrians with less trajectory information.Secondly,the ability to extract information about the interaction between pedestrians is somewhat insufficient,and there has been inadequate feature extraction for related interactions.Finally,existing models have a large amount of redundant computation during training,which not only slows down the training rate but also prevents further improvement in the accuracy of the model prediction.Based on the above problems,This thesis proposes an Autoencoder-based model for pedestrian trajectory prediction of variable length(ASTRAL).The goal of the algorithm research is to extract trajectory features that have a significant impact on predicting future trajectories of pedestrians by using a large amount of historical trajectory information generated from pedestrian activities,to more accurately predict their future activities.This thesis will collect data and process trajectory features for the predicted pedestrian and nearby pedestrians’ trajectories,and then predict more accurate future trajectory information through relevant algorithms involved in the prediction model.The trajectory prediction task is divided into several parts: first is feature extraction for pedestrian historical trajectories,especially for pedestrians with short trajectories; second,appropriate weight allocation algorithms are used to extract interaction information between pedestrians and conduct quantitative analysis of such interaction during pedestrian movements.Finally,further optimization is carried out to improve the overall training efficiency of the model,enabling the prediction model to perform better.At the same time,the related research work of the thesis has undergone multiple strict tests on public datasets to verify the accuracy of the model.Compared with previous experimental results,the prediction results of ASTRAL are more outstanding,which can more effectively predict the future trajectories of pedestrians and play a more important role in many fields.