Nowadays, the issue of predicting the path of human movement in different environments, specifically crowded environments, has become a critical issue for self-driving mobile systems (self-driving cars and social robots) that want to move in human-centered environments, security and surveillance applications. This issue has its own complexities due to the existence of indirect and unwritten social interactions between humans and their physical interactions with objects in the environment. These interactions all originate from the continuity and foresight of pedestrians in determining the future movement path, so it seems that predicting spatial and temporal interactions is vital to prevent pedestrian collisions and accidents. However, most of the methods presented so far have ignored the temporal interactions between pedestrians in the scene. In order to address these challenges, a structure called physical scene coding is proposed in this thesis. This structure presents a hybrid approach based on graph attention network and convolutional neural networks for predicting the movement path of humans. The proposed network is based on a seq2seq (sequence-to-sequence) architecture, which consists of two general parts: Encoder and Decoder. In the proposed method, a composite structure that includes two networks, GAT (graph attention neural network) and CNN (convolutional neural network) is used to extract spatial interactions and an additional layer of LSTM (long short term memory) is used to extract temporal interactions. To eva‎luate the effectiveness of the proposed approach, extensive experiments have been performed on the ETH and UCY datasets, which are among the most common datasets in this field. eva‎luation based on widely used criteria named ADE (average displacement error) and FDE (final displacement error) has been done. The obtained results show that the average error ADE and FDE for the proposed approach of this thesis, it is 0.42 and 1.02, respectively, compared to STGAT, which recorded a value of 0.54 and 1.10, has led to a significant improvement. The obtained results indicate the superiority of the proposed approach over the best methods that have been presented before.

سمانه حسيني

مازيار پالهنگ , امير خورسندي كوهانستاني