A pedestrian trajectory prediction method intergrating spatiotemporal information and social interaction features
With the rapid development of computer technology and deep learning algorithms, automation and intelligent technology have become widely researched areas of interest. Accurate inference and prediction of pedestrian trajectories, as a crucial component in fields such as autonomous driving and advanced driver assistance systems, have been the focus of scholars both domestically and internationally. The basic principle of pedestrian trajectory prediction is to forecast the future position coordinates of pedestrians at fixed time intervals based on their past trajectories, incorporating information such as pedestrian location, self-motion history, pedestrian-environment interaction, and pedestrian-pedestrian interaction. Due to the high dynamics, randomness, and complex interaction with the environment, accurately predicting future pedestrian trajectories has always been challenging. Recent research on pedestrian trajectory prediction models can be broadly categorized into two types: model-driven modeling methods and data-driven modeling methods. Model-driven modeling methods simulate pedestrian self-motion and interactions among pedestrians based on artificially designed energy function models to predict pedestrian trajectories. Conversely, data-driven modeling methods treat pedestrian trajectory prediction as a time series forecasting problem, leveraging the excellent fitting and expressive capabilities of machine learning or deep learning to model the correlation of pedestrian motion sequences through statistical analysis of large datasets. Despite their strengths, existing data-driven modeling methods face challenges in effectively capturing and integrating pedestrian temporal motion characteristics and complex abstract social interaction behaviors among pedestrians. To address these challenges, this paper proposes a pedestrian trajectory prediction method that integrates spatiotemporal information and social interaction features. Firstly, historical pedestrian trajectories are obtained, and a motion trajectory mapping module based on a multi-layer perceptron is used to encode preliminary pedestrian historical trajectory information. Then, based on a combination of long short-term memory (LSTM) networks and feature attention mechanisms, a motion spatiotemporal feature encoding module is designed to explore the temporal dependency of pedestrian self-motion sequences within the observation period and selectively capture the spatiotemporal correlation information of pedestrian self-motion sequences. Furthermore, based on the analysis of the complex interactions between pedestrian self-motion and surrounding pedestrians, a pedestrian social interaction information propagation module based on graph convolutional networks (GCN) is introduced to model the social interaction features among pedestrians within the same scene. Finally, leveraging a multi-modal future trajectory decoding module incorporating the concept of Laplace mixture distribution, the integrated analysis and decoding of pedestrian historical trajectory spatiotemporal correlation information and social interaction features are performed to predict trajectory distributions and capture the uncertainty of future trajectories, resulting in multiple future motion trajectories for pedestrians. The proposed model is qualitatively and quantitatively analyzed using the ETH and UCY public datasets in this paper. Average displacement error (ADE) and final displacement error (FDE) are selected to evaluate the performance of the network model on the ETH and UCY datasets. The experimental results are compared with traditional methods and several current mainstream methods. Compared with the optimal models among them, the proposed model reduces ADE and FDE by 2％ and 5％, respectively.The experimental results demonstrate that the proposed method can robustly and reliably achieve pedestrian trajectory prediction, and extensive comparative experiments also confirm the effectiveness of the proposed approach.
NETL
NSTL
万方数据
