An Intelligence-Based Approach for Prediction of Microscopic Pedestrian Walking Behavior

• Published in: IEEE transactions on intelligent transportation systems 2019-10, Vol.20 (10), p.3964-3980
• Main Authors: Ma, Yi, Lee, Eric Waiming, Hu, Zuoan, Shi, Meng, Yuen, Richard Kowkkit
• Format: Article
• Language: English
• Subjects: artificial intelligence; artificial neural network; Artificial neural networks; Behavior; behavioral learning; Computer simulation; Decision making; Errors; Learning theory; Legged locomotion; Mathematical model; Microscopy; Neural networks; Numerical models; Pedestrian walking behavior; prediction; Predictions; Predictive models; Quantitative analysis; Walking
• ISSN: 1524-9050; 1558-0016
• DOI: 10.1109/TITS.2019.2931892

This paper focuses on locally microscopic pedestrian walking behavior and proposes an intelligent behavioral learning approach for its prediction. In this approach, pedestrian walking behavior was modeled as a special artificial neural network whose input and output layers were used to accommodate a pedestrian's perceived environmental information (e.g., information on the destination, obstacles, and neighbors) and his or her walking behavioral response, respectively. The developed neural network was trained based on a large volume of data samples of real-life pedestrian walking behavior (3813 training samples) to acquire knowledge of pedestrian walking behavior and to develop the ability to predict microscopic pedestrian walking behavior. A quantitative evaluation index, R-squared, was calculated to evaluate the learning performance; the mean was calculated as 0.900, which indicates that the neural network can well capture the underlying decision-making mechanism behind pedestrian walking behavior. The approach was verified by the prediction of microscopic pedestrian walking behavior details in two real-life scenarios. The vector displacement error and the speed error were calculated to evaluate the quality of the prediction; the mean of the vector displacement error (the speed error) in two scenarios were, respectively, calculated as 0.192 (0.116) and 0.226 (0.096), which indicates that the prediction results were acceptable from an engineering perspective. Based on these results, we consider the developed approach to be capable of predicting microscopic pedestrian walking behavior. Moreover, extended application also shows that the proposed approach has promise for simulation of the short-term flow of a pedestrian crowd.