Evaluation of Indoor Human Trajectory Regression Techniques

Abstract This paper is divided into two main parts: The first part focuses on automated feature generation and model discovery. It utilizes the Feature Engineering Automation Tool (FEAT) for automated data processing to enhance model performance. Additionally, gplearn(genetic programming) is employed to automatically generate mathematical expressions for trajectory prediction through the simulation of natural selection and genetic variation. The second part emphasizes nonlinear modeling with neural networks. It compares two different neural network structures, MLP (Multilayer Perceptron) and Kans(Kolmogorov Arnold Networks), using the backpropagation algorithm to fit trajectories by minimizing the loss function weights. This section aims to predict indoor pedestrian trajectories and compare the performance of these two neural network models. In this study, both symbolic regression and neural networks are employed for trajectory prediction. A comparative analysis of these two methods is conducted, followed by a detailed comparison of the prediction performance between MLP and Kans within neural networks. The paper discusses the prediction accuracy, computational efficiency, model interpretability, and applicability of the methods used. The findings provide insights into the strengths and weaknesses of symbolic regression and neural networks in trajectory prediction, highlighting their practical significance and potential applications.