Trajectory Prediction for Multiple Types of Traffic Participants at a Signalized Intersection Based on Potential Field and Heterogeneous Graph Attention Network

The work is to investigate the trajectory prediction for multiple types of traffic participants under a signalized intersection scenario within intelligent connected environments based on potential field (PF) and heterogeneous graph attention network (HGAT), where participants include connected and automated vehicles (CAVs), human vehicles (HVs), cyclists, and pedestrians. A novel method of trajectory prediction termed risk potential field-heterogeneous graph attention network (RPF-HGAT) is proposed based on risk potential field (RPF) modeling and heterogeneous graph encoding. The trajectory prediction model that integrates historical temporal features, dynamic interaction features, and multiscale map features is constructed. First, an RPF model is constructed based on oriented boundary distance (OBD) and a piecewise PF function, characterizing the differential interaction intensities in various directions and representing the potential conflicts among participants. Second, three encoders are constructed, where temporal attention network (TANet) is used to extract historical temporal features, potential field-spatial attention network (PF-SANet) is used to capture dynamic interaction features, and multiscale squeeze-and-excitation network (MS-SENet) is used to encode multiscale map features, respectively. Finally, a decoder that fuses multisource features and reflects type-specific characteristics is established to enable differentiated future trajectory prediction for different types of participants. The proposed model is validated through INTERACTION and nuScenes data sets. The results of trajectory prediction show that the MinADE6 and MinFDE6 can reach to 0.15 m and 0.63 m on the INTERACTION data set, respectively, MinADE10, MinFDE10, and MR2,10 can reach to 1.16 m, 2.14 m, and 0.44 on the nuScenes data set, respectively. MinADE6 and MinFDE6 are reduced by 34.8% and 17.1%, respectively, compared with the methods of heterogeneous edge-enhanced graph attention network (HEAT) on the INTERACTION data set. MinADE10, MinFDE10, and MR2,10 are reduced by 23.2%, 30.3%, and 21.4%, respectively, compared with the methods of Trajectron++ on the nuScenes data set. By using the proposed method, potential conflict can be identified in advance, and the level of traffic safety can be enhanced.