Machine Learning-Based Pedestrian Crossing Decision Models to Increase Pedestrian Safety

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Jan 3 2024

In a paper published in the journal Sensors, researchers highlighted the critical role of artificial intelligence (AI) in advancing road traffic safety, particularly within intelligent cities. Their focus centered on employing machine learning (ML), mainly support vector machine (SVM), to improve predictions regarding pedestrian crossings within intelligent transportation systems.

By leveraging open-source computer vision (OpenCV) image recognition and authentic pedestrian behavior data from Chinese city intersections, the study demonstrated that the SVM model outperformed other ML approaches, showcasing its potential for accurately predicting pedestrian behaviors during crossings and applicability in intelligent traffic simulations.

Pedestrian Safety Enhancement

The focus in current road design heavily favors motor vehicles, amplifying pedestrian vulnerability and sustaining a concerning frequency of pedestrian accidents, as evidenced by the 2021 China Traffic Accident Statistical Yearbook. Despite extensive exploration of vehicle crossing gaps, more studies must examine pedestrian crossing decision-making. Research has concentrated on permissible crossing gaps and pedestrian crossing speeds, yet it needs help accurately capturing real-time pedestrian behaviors and decision-making dynamics. By integrating image recognition and ML, this study aims to address these limitations, offering a refined analysis of factors impacting pedestrian crossings, improving decision models, and enhancing the accuracy of behavior simulations.

ML for Pedestrian Analysis

This study adopts an ML methodology to forecast pedestrian crossing probability and speed during street crossings, divided into distinct stages. Firstly, the previous chapter focuses on collecting data concerning pedestrian crossing features and vehicle driving patterns, culminating in acquiring a comprehensive dataset. Following this, the study establishes four distinct ML models, presenting the experimental results for each model. Researchers analyze and compare these models to determine the most suitable approach for this research while discussing the factors influencing the outcomes.

The initial step involves data preprocessing after data collection. Label encoding is applied to convert the categorical feature 'z' into a numerical representation. In contrast, the categorical column 'p' transforms a binary categorical column for model training purposes. The dataset then undergoes z-score standardization to ensure uniformity and consistency.

After the data preparation phase, the study uses four different ML algorithms—decision trees, the Bayesian algorithm, the backpropagation (BP) neural network, and SVM. The researchers independently employ each algorithm to predict crossing probability and crossing speed.

The decision tree algorithm constructs models using tree structures, simplifying predictions through straightforward decision rules. The dataset is progressively divided into smaller subsets while concurrently constructing associated decision trees comprising decision nodes and leaf nodes representing outcomes. This model distinguishes itself for its straightforwardness and ease of comprehension.

Meanwhile, the Bayesian algorithm, grounded in Bayesian probability theory, excels in handling uncertainty and noise. It is beneficial in scenarios with limited data and can provide explicit uncertainty estimates. This adaptability is advantageous in dynamic environments like pedestrian street crossings.

SVMs are supervised learning algorithms founded on statistical learning theory. They excel in addressing classification and regression problems even with limited sample data. Utilizing kernel functions, SVMs map data from the input space into higher-dimensional feature spaces. This process aims to pinpoint an optimal separating hyperplane that distinguishes data points.

The study selects the radial basis function as the kernel function in this research. Multi-layer perceptrons (MLPs) form a feedforward neural network architecture featuring input, hidden, and output layers. With non-linear activation functions and trained via optimization algorithms, MLPs learn complex mapping relationships between input data and output labels. Their flexibility renders them valuable for diverse machine-learning tasks.

Parameter settings play a crucial role in model performance. K-fold cross-validation is a technique used by researchers to minimize overfitting and modify parameters. Each model undergoes specific parameter adjustments based on cross-validation outcomes to optimize performance. The researchers also conduct sensitivity analysis, identifying the relative sensitivities of specific parameters within the scope of this study without significantly impacting the model's overall performance.

Pedestrian Behavior Analysis in Simulation

The results for predicting crossing probability among different ML models revealed distinct performances. While the decision tree model exhibited mediocre predictive capabilities, the SVM model stood out, showcasing the highest accuracy and Kappa coefficient, indicating its balanced sensitivity and specificity. Statistical tests, including McNemar's, emphasized significant differences between models like the MLP and Naïve Bayes models, showcasing specific model disparities. However, multiple experiments revealed a lack of statistical significance among most model comparisons.

The SVM and MLP models consistently displayed higher area under the curve (AUC) values across different folds, underscoring their better performance. For predicting crossing speed, the SVM model outperformed others, exhibiting the lowest mean squared error (MSE) and root mean squared error (RMSE), along with the highest coefficient of determination (R2) and lower mean absolute deviation (MAD), making it the most suitable model overall.

Feature analysis sheds light on the factors influencing pedestrian crossing decisions. SVM-based models highlighted the significance of vehicle speed and distance in predicting crossing probability. Different age groups exhibited distinct tendencies: while elderly pedestrians prioritized more considerable vehicle distances, children tended to make faster crossings despite varying vehicle speeds and distances. Analysis via SHapley Additive exPlanations (SHAP) plots provided more profound insights into the impact of features on prediction outcomes, emphasizing the importance of 'x' and 'y' features over 'z' across different models. This detailed understanding can contribute to enhancing pedestrian simulation accuracy in traffic software like planung transport verkehr-verkehr in vtadten – simulationsmodell (PTV-VISSIM2020), offering realistic representations of pedestrian behaviors and safety considerations, which are currently limited in existing pedestrian modules.

Conclusion

To summarize, this study utilized ML methods with real-world data from Dalian, China, to predict jaywalking behaviors at intersections. The SVM model showed exceptional accuracy and sensitivity in forecasting pedestrian crossing probabilities. While the findings support pedestrian safety in intelligent vehicle systems, they might only partially translate to other traffic environments. Future research will broaden data collection to include diverse vehicles and urban settings, considering additional factors influencing pedestrian behavior. Plans involve integrating these insights into traffic simulation systems and developing real-time predictive models for jaywalking in intelligent vehicles.