Research On Prediction Method Of Pedestrian Trajectory From The First Perspective

The analysis of pedestrian 'behavior based on the first-view video is a research hotspot in the field of computer vision in recent years,and has attracted widespread attention from domestic and foreign researchers,as an important task in behavior analysis,motion prediction can be applied to many fields related to intelligent transportation such as autonomous driving and service robot navigation.The work of this paper is mainly to analyze the historical movement trajectory of the target pedestrian in the video from the first perspective,according to the given video clip,so as to predict its future position.This paper combines a one-dimensional convolutional neural network and a long and short term memory network,and proposes a model based on a encode-decode structure to model and learn the historical information of pedestrian positions and scales,and camera egomotion information that extracted from video frames.Among them,the future information of the camera's egomotion is used to guide the model to better predict the pedestrian's future trajectory.The long and short-term memory network is used to predict the encoded feature vector,and then the prediction sequence is decoded.In order to evaluate the performance of the proposed prediction model,it was tested on the MOT 16 pedestrian dataset,FPL pedestrian dataset and the FPP pedestrian dataset collected in this paper,and compared with other trajectory prediction related algorithms.The test results show that the pedestrian trajectory prediction model proposed in this paper has a good performance in predicting the pedestrian's future trajectory.The innovation 'of this paper mainly include the following aspects:(1)Aiming at the problem of pedestrian trajectory prediction from the first perspective,because the one-dimensional convolutional neural network can extract the characteristics of the features in the one-dimensional sequence,and long and short-term memory networks can handle the characteristics of time series,a model combining the two neural networks,namely CNN-LSTM pedestrian trajectory prediction model is proposed.(2)The future information of the camera's egomotion is added to the model,with the historical information of the pedestrian position and scale,and the historical information of the camera's egomotion,as key factors to guide the model to make predictions to improve the prediction accuracy of the model.(3)Aiming at the problem of fewer first-view datasets in the current pedestrian trajectory prediction research work,a new pedestrian dataset recorded from the first perspective was collected and marked,which provided the learning data required for modeling for similar research work.