Evaluation Of Pedestrian’s Walking Trajectory Prediction Algorithm Based On Artificial Neural Network

With the rapid development of information age,pedestrian’s positioning and navigation plays an increasingly important role in military,archaeology and daily travel.There are two kinds of methods commonly used for pedestrian’s trajectory estimation.The most common method is based on kinetic model,while the emerging method is based on neural network.Neural network plays an extremely important role in the area of big data process.With the accumulation of data in these decades,neural network has surpassed the traditional machine learning method and is exclusive in many fields,such as image processing,speech recognition,natural language processing,etc.Considering about the algorithm of pedestrian’s walking trajectory prediction based on neural network,it can be divided into two categories.The first category is to estimate the single state variables in pedestrian walking,such as walking speed,moving displacement,turning angle,etc.The second category uses the End-to-End Learning Framework,it can directly product the total trajectory only with the raw input data putting into the neural network.These positioning methods have their own advantages and disadvantages in different application scenarios,so it’s difficult to entirely compare the performance of each method.This paper intends to evaluate the performance of each method in the same application scenarios and evaluation criteria.Because the traditional kinetic model is to divide the steps events by detecting the time point of pedestrian heel landing,so as to get the step length,while the method based on neural network can only get the displacement in a fixed period of time when the window of input data is fixed during training,so this paper intends to compare the accuracy of each method by the total error of the same period of data.Because the acceleration data and angular velocity data collected during pedestrian movement have strong time-series correlation,this paper intends to use a network with strong performance in processing time-series data for comparative experiments.In the training process of neural network,a very important premise is to prepare dataset.There are two difficulties in the preparation process of data set.The first is that the data collection process is easily affected by external noise and internal noise,even the existing data collected by high-precision visual inertial odometer also has noise;The second is that the amount of data is not enough,and the end-to-end structure needs to estimate more information than only estimating the step length or displacement of the network structure,so the end-to-end structure needs more dataset to train.In this paper,we intend to use data enhancement to expand the dataset,including the sliding window method and the Synthetic Minority Oversampling Technique(SMOTE).In order to compare the traditional methods based on dynamic model,single state variable estimation based on neural network and End-to-End Learning Framework,this paper designs a comparative experiment of the device walking in the hand-held state.The experimental results show that in this application scenario,the neural network method is better than the traditional method based on dynamic model,and the end-to-end structure has better performance than the single variable method in trajectory estimation,and the performance after data enhancement is better than that before data processing.This experiment provides a basis for the selection of positioning methods in different scenarios in the future,and it has a great reference value.