Pedestrian Crossing Trajectory Prediction Method Based on Deep Learning

Predicting pedestrian trajectories plays an important role in improving the safety of driverless driving, while pedestrians have extremely high uncertainty. It is difficult to correctly predict trajectories using traditional physics prediction, and deep learning neural networks have more advantages. In order to analyze and improve the prediction of pedestrian crossing trajectory based on deep learning, this paper constructs LSTM network model, TCN network model and Transformer network model, and applies them to a virtual 3D dataset for trajectory prediction, and compares and analyzes the training performance and prediction performance of the three models. Additionally, the Transformer network is enhanced with an autoregressive mechanism, a positional embedding mechanism, and optimized training strategies, resulting in the Transformer model achieving the best performance for pedestrian trajectory prediction on this dataset. This paper summarizes the characteristics of different neural network models in pedestrian trajectory prediction and outlines future research directions.