EGAT: Extended Graph Attention Network for Pedestrian Trajectory Prediction
Algorithm 1 Framework of FUM to model spatial interaction.
Input: the set of observed trajectories for N pedestrians on the current batch, , denoted as X;
Output: global interaction feature for N pedestrians after graph convolution of l layers, denoted as ;
(1)
Embedded Vectors. The pedestrian coordinate is embedded into a fixed length vector , and the set of for N pedestrians is denoted as ;
(2)
TS-LSTM. is used to encode a single pedestrian, and the output is ;
(3)
Feature Updating (FU). H as input, the global interaction feature Z for all nodes is updated by FU, ;
(4)
Graph Construction. In the spatial domain, N pedestrians are regarded as nodes, and the connections between pedestrians are regarded as edges to construct a graph. The graph structure is represented by adjacency matrix , is the adjacency matrix with self-connection;
(5)
Attention Calculation. The attention coefficient is calculated by and at each time step t, and the set of constitutes the attention matrix ;
(6)
Graph Convolution. Features of pedestrians are aggregated by graph convolution, ,, is a weight matrix, is softmax operation;
