Predicting the movement path of humans is a critical task in various applications, including autonomous vehicles, social robotics, and crowd management. Various methods have been presented in this field, including the Social LSTM method. The Social LSTM model has been widely used to predict the movement path of humans, but it has limitations, such as treating all neighbors the same regardless of their importance. To address this issue, in this thesis, an innovative movement path prediction approach called Attention Social LSTM is proposed, which uses the attention mechanism to selectively focus on relevant interactions during human movement path prediction. In the proposed method, attention coefficients are first extracted using a network based on reinforcement learning. These attention coefficients determine the importance of each neighbor when predicting the movement path of a target human. By incorporating speed information into the attention mechanism, the proposed model gains a deeper insight into the dynamics of human movement, leading to more accurate predictions. These attention coefficients are extracted by the SARL method. The presented Attention Social LSTM architecture sends attention coefficients, position data, and social pooling module information to long-short-term memory blocks and provides a comprehensive framework for human movement path prediction. The attention mechanism significantly improves the model’s ability to capture meaningful interactions, leading to better prediction of the performance of humans’ movement paths. To eva‎luate the effectiveness of the proposed approach, extensive experiments have been conducted on ETH and UCY datasets. The eva‎luation is based on the widely used average displacement error (ADE) and final displacement error (FDE) criteria. The results show the superiority of the proposed approach of this thesis compared to the Social LSTM model. On the ETH and UCY datasets, the proposed approach achieves an average reduction of 6.2% in ADE error and 6.3% in FDE error compared to the Social LSTM method. The proposed approach of this thesis provides a significant improvement over the Social LSTM model for predicting human paths. By using the attention mechanism, the proposed model can effectively focus on human interactions, which leads to more accurate and reliable trajectory predictions.

سمانه حسيني

مهران صفاياني , مازيار پالهنگ