Pedestrian safety at unmarked crosswalks is a significant concern in urban transportation. With the advancement of self-driving technologies like autonomous robots and self-driving cars, accurately predicting pedestrian intentions along their pathways, including crossings and designated paths, becomes imperative. Recent strides in predictive models rooted in deep learning algorithms have facilitated the development of predictive systems for pedestrian crossings. In this thesis, we explore the feasibility of forecasting pedestrian crossing behavior at unmarked crosswalks employing the JAAD dataset and models that leverage both visual data, such as video frames, and non-visual data, such as the coordinates of key points on the pedestrian's body. The JAAD dataset, a comprehensive video dataset comprising annotated pedestrian and vehicle trajectories at unregulated crossings, serves as the foundation for training and eva‎luating the models devised in this study. This dataset encompasses environmental variables, pedestrian attributes, encompassing the coordinates of key points on the pedestrian's body and bounding box, as well as traffic conditions, including vehicle speed. The models developed in this research are subjected to diverse performance metrics for predicting pedestrian crossings and are benchmarked against other models in the field. These models fall into two primary categories: pedestrian crossing event prediction models relying on a fusion of non-visual features and pedestrian crossing event prediction models predicated on the prediction of pedestrian movement paths, utilizing visual data. Models grounded in non-visual features predict pedestrian crossings through the amalgamation and processing of non-visual attributes, such as key points on the pedestrian's body, bounding box coordinates, and vehicle speed. Meanwhile, the path-based model, known as BaseGAN, processes a sequence of input frames to extract crucial features. Subsequently, it generates the distribution of the pedestrian's movement path in future frames and predicts pedestrian crossings based on this generated path. The results affirm the superiority of the designed models in terms of predictive accuracy. Notably, the BaseGAN model outperforms all other pedestrian crossing event prediction methods, achieving an F1 score of 83% on the JAAD dataset. This research concludes that the BaseGAN deep learning model holds potential for enhancing pedestrian safety by issuing early warnings to drivers, thus mitigating potential collisions with pedestrians. The thesis also delves into the research's limitations and presents avenues for future investigations. The findings gleaned from this research offer valuable insights into advancing self-driving car safety and contribute to shaping the future of autonomous transportation while pioneering innovative approaches to pedestrian safety.

سعيد صدري , بهزاد نظري

محمدرضا احمدزاده , نادر كريمي