A new hybrid information fusion method for trajectory prediction

Pedestrian trajectory prediction has numerous applications in various fields, such as autonomous driving, advanced video surveillance, etc. The performance of trajectory prediction can be improved in various ways, such as fully utilizing multi-source information, using more effective networks, considering more complex social interactions, etc. The multiple sources of information interact with each other as an entire system, so considering only the local impacts between pairs is not sufficient. Because in real life it is difficult to isolate pairwise interactions between multiple sources of information and consider them independently. Hybrid information fusion can effectively integrate multi-source information. In this paper, a new hybrid information fusion approach is proposed that takes into account both local and global influences among multi-source information. First, multi-source information is obtained, including historical trajectory information, social interaction information and scene interaction information. Second, the local and global impacts between multi-source information are independently modeled. Third, the multi-source information is fused by the power average (PA) operator. Finally, the result from the PA fusion, social interaction information and scene interaction information are fused through concatenation fusion method. As a result, multi-source information is fully utilized and information loss is reduced. Experimental results demonstrate that the proposed model achieves greater accuracy in parts of the ETH/UCY datasets compared with some baselines.