Attention-addressing and adaptive-intention-clustering based memory recall for pedestrian trajectory prediction

Pedestrian trajectory prediction is widely applied in autonomous driving, service robots, surveillance systems, etc. The trajectory prediction method using the memory of the retrospective visible examples is more explanatory. But compressing instance information can cause feature loss, which will lead to poor prediction accuracy. Thus, the Attention-addressing and Adaptive-intention-clustering based Memory Recall (AAMR) for intention prediction is proposed in this paper. AAMR introduces a stable attention mechanism to assign weights when addressing similar instances in the attention addressor. It filters out input features that are excessively distant from the corresponding stored instances, thereby compensating for the feature loss caused by sample information compression. AAMR uses a training loss calculation method better suited to training conditions. Additionally, an adaptive-intention-clustering method is proposed to adjust and generate the final predicted intention from multiple addressed intentions. The prediction intention generated by AAMR combines relevant past trajectory information to complete trajectory prediction. AAMR improves ADE and FDE by 4.7 and 8.5% respectively from the previous best method on the ETH/UCY dataset. And ADE and FDE improve by 12.5 and 16.2% respectively on the UNIV dataset.