Pedestrian Trajectory Prediction Based on Transfer Learning for Human-Following Mobile Robots

Recent developments in the field of service robots have led to a renewed interest in human-robot coexistence environments such as at home and office. In this regard, this study focuses on one of such service robots, a human-following mobile robot. In particular, we consider predicting the future trajectory of pedestrians using a machine learning algorithm to improve the accuracy of tracking people. Massive trajectory data is required in existing methods to train the prediction model; however, collecting a sufficient amount of data in general public places before providing services is challenging. Therefore, in this study, we propose a trajectory prediction method based on extracting similar datasets from a large-size dataset and generating a pre-trained prediction model using the extracting datasets. We express the data features in the source and target environments as probability distributions and evaluate the divergence between them. Specifically, the dataset features are expressed as a multidimensional Gaussian distribution and discrete distribution of samples. Then, similarities using the Kullback-Leibler divergence are compared. To verify the effectiveness of the proposed method, we compare the prediction results of the LSTM-based algorithm with those obtained by extracting multiple source datasets from a large dataset and training prediction models using these datasets. The result shows that the proposed method makes it possible to construct an appropriate prediction model with high accuracy in trajectory prediction.