Graph Neural Network for Human Trajectory Forecasting
Predicting human travel trajectories in complex dynamic environments play a critical role in various fields such as autonomous vehicles, intelligent robots, intelligent transportation systems, and also for wireless communication networks. In wireless communication, an accurate user trajectory forecast is crucial for resource allocation, energy efficiency, and quality of service improvement. Due to the high nonlinearity and the complexity of the user's walking behavior, traditional methods cannot satisfy the requirements of mid-and-long term prediction and often ignore spatial and temporal dependencies of each position. In this paper, we propose a novel deep learning framework approach based on Graph Neural networks and long short-term memory, to tackle the prediction problem in realistic human movement. Instead of applying regular convolutional and recurrent units, we formulate the problem on graphs and build the model with complete convolutional structures, which enable faster training with fewer parameters. Specifically, we are interested in predicting the future positions of the user motion given a history of the position for a collection of users trajectories. To train our model we are using the Edinburgh Informatics Forum Pedestrian Database (EIFPD). We apply data mining preprocessing techniques to achieve results that demonstrate where our approach outperforms previous work and show that our model effectively captures comprehensive spatio-temporal correlations.