Pedestrian trajectory prediction method based on the Social-LSTM model for vehicle collision

Journal article, 2024

Techniques for predicting the trajectory of vulnerable road users are important to the development of perception systems for autonomous vehicles to avoid accidents. The most effective trajectory prediction methods, such as Social-LSTM, are often used to predict pedestrian trajectories in normal passage scenarios. However, they can produce unsatisfactory prediction results and data redundancy, as well as difficulties in predicting trajectories using pixel-based coordinate systems in collision avoidance systems. There is also a lack of validations using real vehicle-to-pedestrian collisions. To address these issues, some insightful approaches to improve the trajectory prediction scheme of Social-LSTM were proposed, such methods included transforming pedestrian trajectory coordinates and converting image coordinates to world coordinates. The YOLOv5 detection model was introduced to reduce target loss and improve prediction accuracy. The DeepSORT algorithm was employed to reduce the number of target transformations in the tracking model. Image Perspective Transformation (IPT) and Direct Linear Transformation (DLT) theories were combined to transform the coordinates to world coordinates, identifying the collision location where the accident could occur. The performance of the proposed method was validated by training tests using MS COCO (Microsoft Common Objects in Context) and ETH/UCY datasets. The results showed that the target detection accuracy was more than 90% and the prediction loss tends to decrease with increasing training steps, with the final loss value less than 1%. The reliability and effectiveness of the improved method were demonstrated by benchmarking system performance to two video recordings of real pedestrian accidents with different lighting conditions.