How Can a Drone and AI Assess Pedestrian Risk at Intersections?

Cities are expanding rapidly, yet pedestrian safety remains a challenge. A research team has devised a new way to gauge crossing risk using only drone‑captured video. The footage is processed by YOLOv8, an object‑detection algorithm, and pedestrian paths are tracked with ByteTrack, yielding frame‑by‑frame trajectories without ground sensors.

They defined a "behavior spectrum" based on five features: speed, acceleration, crossing time, remaining green‑light time, and red‑phase duration. Values are normalized using the interquartile range method. Then, an enhanced CRITIC technique assigns objective weights to each feature, considering median variation and absolute correlation.

The resulting risk eigenvalues are clustered with K‑means into four levels: no risk, low, medium, high. The study examined 1,210 crossings at an eight‑lane intersection in Harbin, China. Of the pedestrians, 576 complied with the signal, 634 did not. Non‑compliant walkers showed a mean speed of 1.29 m/s, more variable than compliant ones (1.40 m/s). Groups were distinguished with silhouette scores of 0.60 and 0.69. An expert validated the classification on 20 samples, achieving a Fleiss' Kappa of 0.87.

This method provides a systematic, transparent assessment of pedestrian risk, valuable for both research and urban safety management.

The proposed method is grounded in Behavior Spectrum Theory (BST), which treats pedestrian behavior as a multidimensional vector. The five selected variables represent the main axes of this space: speed (v) and acceleration (a) describe movement dynamics; crossing time (t_cross) indicates the duration of interaction with the intersection; remaining green time (t_green) measures the safe window for crossing; red‑phase duration (t_red) adds contextual risk.

Normalization via the interquartile range (IQR) mitigates outlier influence, ensuring that asymmetric distributions do not skew comparisons across variables with different scales. The enhanced CRITIC algorithm computes a median‑based coefficient of variation (CV_med) to assess relative dispersion and uses absolute correlation to quantify inter‑criterion conflict. The weight w_i for each feature i is defined as w_i = CV_med_i * (1 - |ρ_ij|), where ρ_ij is the correlation between i and j. This assigns higher importance to highly variable, weakly correlated variables, reducing informational redundancy.

Risk eigenvalues, derived as weighted linear combinations of the five features, serve as scalar danger indicators. Subsequent K‑means clustering with k=4 leverages minimization of intra‑cluster squared distances, yielding coherent groups. Silhouette coefficients (0.60‑0.69) indicate good separation, though not perfect, suggesting that additional features (e.g., heading direction, traffic density) could refine classification.

Expert validation, measured with Fleiss' Kappa (0.87), confirms high inter‑rater reliability, exceeding the 0.75 threshold commonly deemed "excellent." Practically, the framework can be embedded in urban traffic management systems, providing real‑time alerts for high‑risk pedestrian zones and guiding targeted infrastructure interventions.