U-EASWS

Dataset Description

U-EASWS (Urban Expressway Type-A Short Weaving Segment UAV Vehicle Trajectory Dataset Integrated with Interaction and Risk Semantic Annotations) is a UAV-based microscopic vehicle trajectory dataset specifically developed for urban expressway Type-A short weaving segments.

The dataset contains 19,122 vehicle trajectories collected from seven real-world Type-A short weaving scenarios in Changchun, China, with approximately 3.6 h of segment-level observation coverage when the seven scenario durations are summed. The approximately 3.6 h value represents the sum of scenario-level observation durations, with the two directions of shared bidirectional recordings counted separately. The observed weaving segments cover different roadway geometries and channelization configurations, including unrestricted weaving, restricted merging, and restricted diverging layouts.

Type-A weaving is characterized by merging and diverging traffic streams that are each required to perform a lane change within the same weaving area. In short weaving segments, the limited longitudinal distance available for completing these maneuvers produces frequent mandatory lane-changing interactions among merging, diverging, and through vehicles. U-EASWS was therefore designed specifically to provide continuous microscopic observations of vehicle movements under these geometrically constrained and interaction-intensive conditions.

In addition to reconstructed vehicle trajectories and kinematic states, U-EASWS provides road-geometry information, lane information, surrogate safety measures, and vehicle-interaction semantic annotations. Frame-level detection and tracking outputs are also released to allow users to perform independent coordinate transformation, trajectory reconstruction, smoothing, vehicle-dimension estimation, or other customized post-processing procedures.

The dataset is licensed under CC BY 4.0.

Supported Research Tasks

U-EASWS is intended to support research on microscopic traffic flow and driving behavior in complex weaving environments, including:

• microscopic driving behavior modeling and model validation;
• car-following, lane-changing, and weaving behavior analysis;
• discretionary and mandatory lane-changing analysis;
• traffic surrogate safety assessment and conflict identification;
• multi-vehicle interaction analysis;
• spatial interaction and driving-behavior research;
• vehicle trajectory prediction;
• traffic simulation model calibration and validation;
• autonomous-driving simulation;
• tactical decision-making and trajectory-planning research;
• comparative analysis of traffic behavior under different weaving geometries and channelization configurations.

Dataset Repository Structure

The repository is organized according to the processing level and semantic content of the data. The current release contains five data/documentation components and a principal processing-code folder.

1. Weaving Segment Geometry Images

Weaving_Segment_Geometry_Images

This folder contains four top-down aerial geometry images covering seven weaving segments:

• WS1_2.png: WS1 and WS2.
• WS3.png: WS3.
• WS4_5.png: WS4 and WS5.
• WS6_7.png: WS6 and WS7.

The images provide the roadway geometry and channelization context required to interpret the trajectory data. They show the actual road layout, lane arrangement, lane boundaries, solid and dashed lane markings, and the geometric organization of the observed Type-A short weaving segments.

The seven scenarios contain multiple channelization configurations, including unrestricted weaving, restricted merging, and restricted diverging layouts. Dashed lane markings indicate roadway sections where lateral crossing is permitted, whereas solid lane markings represent sections with restricted lane-changing opportunities.

These geometry images should be used together with the trajectory fields Direction, Lane number, Dist_to_Merge_m, and Dist_to_Diverge_m when reconstructing the spatial relationship between vehicle movements and the underlying road geometry.

2. Raw Data

Raw_Data

The Raw_Data folder contains frame-level vehicle detection and tracking outputs before coordinate transformation, trajectory reconstruction, kinematic calculation, and semantic annotation.

The raw data are divided into 16 CSV files:

• WS1_2_raw_detection_data_1.csv
• WS1_2_raw_detection_data_2.csv
• WS1_2_raw_detection_data_3.csv
• WS1_2_raw_detection_data_4.csv
• WS1_2_raw_detection_data_5.csv
• WS1_2_raw_detection_data_6.csv
• WS3_raw_detection_data_1.csv
• WS3_raw_detection_data_2.csv
• WS3_raw_detection_data_3.csv
• WS4_5_raw_detection_data_1.csv
• WS4_5_raw_detection_data_2.csv
• WS4_5_raw_detection_data_3.csv
• WS4_5_raw_detection_data_4.csv
• WS6_7_raw_detection_data_1.csv
• WS6_7_raw_detection_data_2.csv
• WS6_7_raw_detection_data_3.csv

Files beginning with WS1_2, WS4_5, and WS6_7 correspond to the three bidirectional observation sites, whereas files beginning with WS3 correspond to the scenario retaining one traffic direction.

Each row represents one tracked vehicle observation in a particular video frame.

The released raw data contain the following fields:

• Frame: video frame index. Unit: dimensionless frame index.
• Vehicle_ID: vehicle tracking ID assigned during the detection and multi-object tracking stage. Unit: dimensionless identifier.
• x_center_px: horizontal pixel coordinate of the center of the detected vehicle bounding box. Unit: pixel.
• y_center_px: vertical pixel coordinate of the center of the detected vehicle bounding box. Unit: pixel.
• width_px: width of the axis-aligned vehicle detection bounding box. Unit: pixel.
• height_px: height of the axis-aligned vehicle detection bounding box. Unit: pixel.

The raw files preserve image-space vehicle positions and detection-box dimensions and can therefore be used for independent coordinate transformation, vehicle-dimension estimation, trajectory interpolation, smoothing, or other user-defined post-processing procedures.

The Vehicle_ID in the raw data is the tracking identifier associated with the corresponding source sequence, while Ego_Vehicle_ID identifies vehicles in the processed trajectory files. When combining observation groups, file/source identity should be used together with the corresponding vehicle identifier.

In this repository, "raw data" refers to the frame-level vehicle detection and tracking outputs used as inputs for subsequent trajectory processing.

3. Basic Trajectory Data

Basic_Trajectory_Data

This folder contains four CSV files corresponding to the four observation groups:

• Weaving Segments 1–2;
• Weaving Segment 3;
• Weaving Segments 4–5;
• Weaving Segments 6–7.

The files provide reconstructed vehicle trajectories expressed in a road-aligned variant of the Frenet coordinate system together with vehicle kinematic states, lane information, traffic direction, geometry-related distances, and estimated vehicle length.

The main fields are defined as follows:

• Ego_Vehicle_ID: processed vehicle identifier interpreted together with the corresponding observation-group identity. Unit: dimensionless identifier.
• Frame: frame index in the corresponding processed sequence; source-to-processed mappings are used when linking raw and processed records. Unit: dimensionless frame index.
• Delta_t_s: elapsed time from the previous record of the same vehicle after ordering by frame; the first trajectory record serves as the temporal reference and uses a blank value by definition. Unit: s.
• x_center: longitudinal vehicle-center coordinate in the road-aligned variant of the Frenet coordinate system. Unit: m.
• y_center: lateral vehicle-center coordinate in the road-aligned variant of the Frenet coordinate system. Unit: m.
• v_x: longitudinal vehicle velocity. Unit: m/s.
• v_y: lateral vehicle velocity. Unit: m/s.
• a_x: longitudinal vehicle acceleration. Unit: m/s².
• a_y: lateral vehicle acceleration. Unit: m/s².
• Direction: vehicle travel direction. The values are Left or Right. Unit: dimensionless categorical variable.
• Lane number: current lane occupied by the vehicle. Unit: dimensionless lane index.
• Dist_to_Merge_m: absolute longitudinal distance from the current vehicle center to the predefined merge reference cross-section. Unit: m.
• Dist_to_Diverge_m: absolute longitudinal distance from the current vehicle center to the predefined diverge reference cross-section. Unit: m.
• Vehicle_Length_m: estimated physical length of the vehicle. Unit: m.

Vehicle_Length_m is estimated as a vehicle-level physical attribute from multiple valid frame-level bounding-box observations and remains constant for the complete trajectory of the same vehicle within an observation group.

Dist_to_Merge_m and Dist_to_Diverge_m explicitly connect individual vehicle states to the principal geometric locations of the weaving segment and can be used to analyze how microscopic behavior changes with the absolute distance to the merge or diverge reference cross-section.

4. Risk Semantic Data

Risk_Semantic_Data

This folder contains four CSV files with risk semantic annotations corresponding to:

• Weaving Segments 1–2;
• Weaving Segment 3;
• Weaving Segments 4–5;
• Weaving Segments 6–7.

The risk semantic files retain the basic trajectory and kinematic fields and additionally provide surrogate safety measures and explicit risk-related annotations.

The additional fields are:

• TTC: Time-to-Collision. Unit: s.
• PET: Post-Encroachment Time. Unit: s.
• DRAC: Deceleration Rate to Avoid a Crash. Unit: m/s².
• Max_Risk_Target_ID: ID of the target vehicle associated with the conflict risk for the ego vehicle in the current frame; -1 is used as the default target identifier. Unit: dimensionless identifier.
• Risk_Level: explicit risk classification, with values High Risk or Low Risk.

TTC and DRAC are frame-wise state-based surrogate safety measures. PET is treated as an event-level measure associated with identified cut-in events. For each identified cut-in event, the calculated PET value is assigned at the event-detection frame to both the cut-in vehicle and its corresponding following vehicle.

A record is labeled High Risk when at least one of the following conditions is satisfied:

• TTC < 3.0 s;
• DRAC > 3.0 m/s²;
• the record corresponds to an identified cut-in event with PET < 2.5 s.

Records satisfying none of these conditions are labeled Low Risk.

5. Interaction Semantic Data

Interaction_Semantic_Data

This folder contains four CSV files with vehicle-interaction semantic annotations corresponding to:

• Weaving Segments 1–2;
• Weaving Segment 3;
• Weaving Segments 4–5;
• Weaving Segments 6–7.

The interaction semantic files retain the basic trajectory information and add vehicle-pair interaction information derived from a two-dimensional Proximity Resistance (PR) framework.

For each ego vehicle, surrounding vehicles traveling in the same direction within a local perception radius of 50 m are first treated as candidate interaction vehicles. Candidate vehicle pairs are then represented in an ego-vehicle-centered local coordinate system and assigned interaction-related variables.

The additional fields are:

• Interaction_Target_ID: ID of the target vehicle paired with the ego vehicle. Unit: dimensionless identifier.
• x_local: relative-position component of the target vehicle along the local X-axis of the vehicle-pair coordinate system. Unit: m.
• y_local: relative-position component of the target vehicle along the local Y-axis of the vehicle-pair coordinate system. Unit: m.
• v_rel: scalar relative speed between the vehicle pair. Unit: m/s.
• Interaction Strength: dimensionless PR-based measure of proximity-induced vehicle interaction, ranging from 0 to 1. Larger values represent stronger proximity-induced interaction.
• Is Interacting: binary indicator of whether the candidate vehicle pair satisfies the effective-interaction criterion. 1 denotes PR greater than exp(-1) and 0 otherwise. Candidate pairs below this threshold remain in the annotated data.
• Interaction Type: behavioral interaction category, including Car Following, Parallel Driving, and Lane Changing/Weaving. Records in the default no-candidate state are encoded as Safe in the annotation output.

Interaction Strength quantifies proximity-induced interaction, while TTC, PET, and DRAC characterize conflict-related safety states. A large PR value indicates that the target vehicle lies deeper within the empirically inferred driver space of the ego vehicle and therefore represents stronger spatial interaction.

Consequently, Is Interacting = 1 identifies an effective proximity-induced interaction, while the risk semantic fields characterize conflict-related states. The two semantic layers provide complementary dimensions of microscopic traffic behavior.

6. Principal Processing Code

Principal Processing Code

This folder contains the 11 principal processing and supporting files described in the reprocessing guide below. The code README provides detailed inputs, parameter settings, and implementation details.

Road Geometry and Scenario Information

The dataset covers seven Type-A short weaving segments located on urban expressways in Changchun, China.

All seven sites exhibit the defining Type-A weaving configuration, in which merging and diverging traffic streams must each perform a lane-changing maneuver within the same weaving area.

Three UAV observation sequences contain bidirectional traffic flows and each corresponds to two weaving scenarios:

• WS1 and WS2;
• WS4 and WS5;
• WS6 and WS7.

The remaining observation group corresponds to WS3 and retains one traffic direction after preprocessing.

WS1, WS3, WS4, and WS6 retain left-to-right vehicle movement in the road-geometry representation, whereas WS2, WS5, and WS7 represent the opposite right-to-left traffic direction.

The seven scenarios have the following collection characteristics:

• WS1: collected on January 7, 2025; 1547 s of valid video; approximately 290 m observation length; approximately 220 m UAV altitude; four lanes; 80 km/h speed limit; 60 fps source video; 2056 trajectories.
• WS2: collected on January 7, 2025; 1547 s of valid video; approximately 290 m observation length; approximately 220 m UAV altitude; four lanes; 80 km/h speed limit; 60 fps source video; 1935 trajectories.
• WS3: collected on September 17, 2025; 2394 s of valid video; approximately 315 m observation length; approximately 270 m UAV altitude; three lanes; 80 km/h speed limit; source observations include 60 fps and 30 fps video; 2985 trajectories.
• WS4: collected on September 16, 2025; 1437 s of valid video; approximately 375 m observation length; approximately 360 m UAV altitude; four lanes; 80 km/h speed limit; 30 fps source video; 2017 trajectories.
• WS5: collected on September 16, 2025; 1437 s of valid video; approximately 375 m observation length; approximately 360 m UAV altitude; five lanes; 80 km/h speed limit; 30 fps source video; 2126 trajectories.
• WS6: collected on September 17, 2025; 2306 s of valid video; approximately 370 m observation length; approximately 350 m UAV altitude; five lanes; 80 km/h speed limit; 30 fps source video; 4038 trajectories.
• WS7: collected on September 17, 2025; 2306 s of valid video; approximately 370 m observation length; approximately 350 m UAV altitude; five lanes; 80 km/h speed limit; 30 fps source video; 3965 trajectories.

The reported observation length represents the retained roadway coverage used for trajectory extraction, including the core weaving area and adjacent upstream and downstream buffer regions where applicable.

The geometry images in Weaving_Segment_Geometry_Images should be consulted when performing lane-level or geometry-specific analyses because lane numbers, solid/dashed markings, and merge/diverge locations vary among the seven scenarios.

Coordinate Systems

U-EASWS contains three distinct coordinate systems corresponding to different processing levels, with each coordinate system serving a specific role in the processing workflow.

Source-Image Pixel Coordinate System

The frame-level files in Raw_Data use the source-video image coordinate system.

The coordinate origin is located at the upper-left corner of the video image. The horizontal pixel coordinate increases from left to right, and the vertical pixel coordinate increases from top to bottom.

The relevant variables are:

• x_center_px: horizontal bounding-box-center coordinate, in pixels;
• y_center_px: vertical bounding-box-center coordinate, in pixels;
• width_px: bounding-box width, in pixels;
• height_px: bounding-box height, in pixels.

These variables describe image-space measurements, while metric physical distances are provided after coordinate transformation.

Road-Aligned Variant of the Frenet Coordinate System

The preprocessed basic trajectory data adopt a road-aligned variant of the Frenet coordinate system, where the longitudinal and lateral coordinates describe the along-road and cross-road directions, respectively.

For scenes recorded by multiple UAV video sequences, a stabilized reference frame is selected as the master frame. Static road features visible in both the source frames and the master frame are used as control points, and fixed planar homography transformations are applied to project vehicle locations from different UAV views into a common reference coordinate system.

Static road reference points with known physical distances are then used to determine the metric scale. A local coordinate origin is defined in the master frame.

In this coordinate system:

• the x axis represents the longitudinal road direction;
• the y axis represents the lateral road direction;
• positions are measured in meters;
• velocities are measured in meters per second;
• accelerations are measured in meters per second squared.

The principal variables expressed in this system are x_center, y_center, v_x, v_y, a_x, and a_y.

For files containing bidirectional traffic, the same road-aligned variant of the Frenet coordinate system is used for both travel directions, while the Direction field identifies the corresponding vehicle travel direction.

Vehicle-Pair Local Interaction Coordinate System

The interaction semantic data additionally use an ego-vehicle-centered local coordinate system for each candidate vehicle pair.

The centroid of the ego vehicle (EV) is defined as the origin.

The direction of the relative-velocity vector between the ego vehicle and the principal other vehicle (POV) is defined as the positive local Y-axis. The direction perpendicular to the relative-velocity vector and pointing to the right is defined as the positive local X-axis.

The relative position of the target vehicle is projected into this local coordinate system to obtain x_local and y_local.

The vehicle-pair local coordinate transformation is defined for pairs with nonzero relative velocity, for which the relative-velocity direction provides the local-axis orientation.

This vehicle-pair local coordinate system complements the road-aligned variant of the Frenet coordinate system used for x_center and y_center by describing pairwise relative positions.

Temporal Resolution

The UAV videos were recorded using two source recording modes:

• 3840 × 2160 pixels at 30 frames per second;
• 2688 × 1512 pixels at 60 frames per second.

Therefore, the nominal physical interval between two consecutive source frames is:

• 1/30 s, approximately 0.03333 s, for 30 fps video;
• 1/60 s, approximately 0.01667 s, for 60 fps video.

U-EASWS contains observations originating from both source frame rates, and Delta_t_s provides the corresponding physical time interval in the processed trajectory data.

The processed trajectory files therefore include Delta_t_s, which explicitly provides the physical time interval associated with the corresponding trajectory record.

Within a vehicle trajectory, sort by Frame and accumulate valid Delta_t_s values to obtain elapsed time relative to the first observation. The first record serves as the temporal reference. For interaction files, the ego-vehicle time axis can be constructed from unique ego/frame records.

Source-specific frame-rate information should be used together with the corresponding raw records during reprocessing.

Data Collection

The original vehicle observations were extracted from UAV aerial videos recorded using two DJI Mavic 2 Zoom UAVs.

The recording modes were 3840 × 2160 at 30 fps and 2688 × 1512 at 60 fps. The UAVs were operated from a near-vertical top-down perspective to provide continuous coverage of the short weaving areas and their adjacent roadway sections.

A fixed zoom setting was maintained during data collection to preserve a consistent image scale and geometric configuration throughout each recording.

The UAV recording altitude ranged approximately from 220 m to 360 m, while the retained observation lengths ranged approximately from 290 m to 375 m.

Data collection was performed during daytime periods under clear and low-wind conditions to support stable UAV observations and consistent image measurements.

Data Processing

The processed trajectory and semantic data were generated through a standardized workflow.

The major stages include UAV video stabilization and region-of-interest extraction, vehicle detection and multi-object tracking, lane-region annotation, multi-video coordinate registration, transformation from image coordinates to road-aligned Frenet-based metric coordinates, trajectory quality screening, trajectory-gap reconstruction, trajectory smoothing, vehicle kinematic calculation, vehicle-length estimation, surrogate-safety-measure calculation, and interaction semantic annotation.

Vehicle trajectories are reconstructed using PCHIP interpolation where sufficient observations are available, with linear interpolation applied to sparse cases. Symmetric exponential moving average (sEMA) filtering is subsequently applied to improve trajectory smoothness while reducing temporal phase shift.

The processed trajectory and semantic data complement the frame-level records provided in Raw_Data by representing subsequent processing stages.

Principal Processing Code and Reprocessing Guide

The Principal Processing Code folder contains 11 files. These are executable Python processing modules, supporting utilities, a calibration template, dependency requirements, and documentation.

Files and Functions

• coordinate_transform.py registers source-image coordinates to a master frame using a planar homography and converts pixel coordinates into road-aligned Frenet-based metric coordinates. It requires a completed calibration JSON and the corresponding source identifier. It retains source identity and adds metric vehicle-center coordinates.
• filter_trajectory_outliers.py screens metric trajectories using observed duration, observation count, spatial extent, and average path speed. It produces retained records, a separate screening-output file, and a per-vehicle screening summary. The same screening is also integrated into reconstruction, so separate execution is optional.
• reconstruct_trajectories.py screens trajectories, reconstructs trajectory gaps, interpolates positions, fills lane/direction attributes, applies symmetric exponential moving average (sEMA) smoothing, and calculates velocities and accelerations. It also preserves source/original identifiers and produces timing information. Measured merge and diverge reference coordinates can optionally be supplied to generate the corresponding distance fields.
• estimate_vehicle_length.py estimates a fixed physical length for each vehicle directly from the six-column raw detection records and the calibration JSON. It returns vehicle-level length estimates, the number of retained observations, and quality indicators.
• merge_vehicle_lengths.py joins vehicle-level length estimates to reconstructed trajectories using source identity and original vehicle identity. When lineage fields are provided separately, a verified mapping file is used to connect the corresponding records.
• annotate_risk.py calculates nearest-leader TTC and DRAC, event-level cut-in PET, target identifiers, and risk labels using physical trajectories, time information, lane/direction annotations, and vehicle lengths.
• annotate_interactions.py calculates PR-based interaction strength, the effective-interaction flag, and interaction type from basic trajectories, precomputed candidate-pair states, and calibrated PR parameters. All supplied candidate pairs are retained together with their calculated PR values and interaction flags.
• _common.py provides shared CSV validation and output-writing utilities and is used together with the processing scripts as a supporting module.
• calibration_template.json provides the calibration structure for organizing scene-specific information used in coordinate transformation and vehicle-length estimation.
• requirements.txt lists the processing dependencies and their versions.
• README.md documents detailed input schemas, commands, parameter choices, and implementation details for the processing package. It is separate from this repository-level Dataset Card.

Installation

Download the 11 files together, retaining their published filenames and their common directory. Python 3.12 was used for package testing. From the local repository directory, run:

cd "Principal Processing Code"
python -m pip install -r requirements.txt
python reconstruct_trajectories.py --help

Each processing script provides --help. The commands below use descriptive placeholders; replace them with actual local input/output filenames, the source key defined in your calibration JSON, and the numerical source frame rate. Output filenames are specified explicitly so that the original input files remain unchanged during processing.

Inputs to Prepare

1. Raw detections: one source recording with Frame, Vehicle_ID, x_center_px, y_center_px, width_px, height_px. Pixel coordinates must refer to the same image view used for calibration.
2. Scene calibration: a completed copy of calibration_template.json, including the master source, master-frame origin, reference segments with measured physical distances, and corresponding control points for non-master sources.
3. Lane and direction annotations: verified Lane number and Direction fields are incorporated with the metric trajectories to support lane-based processing and road-geometry interpretation.
4. Recording metadata: the actual constant frame rate for each source, and explicit source/ID/time mappings when combining recordings. Optional merge/diverge distances require measured reference cross-sections in the same road-aligned Frenet-based metric coordinate system.
5. Interaction inputs: interaction annotation uses a candidate-pair CSV and calibrated PR parameter CSV prepared according to the input schemas described below.

For calibration, replace the required null values with measured values and change template_only to false only after completing the configuration. At least four valid corresponding points are required for each non-master homography. The current implementation is configured for horizontally aligned road scenes with rotation_rad = 0, consistent with the coordinate configuration used for U-EASWS. The master-frame scale is the mean of the measured physical-to-pixel distance ratios. Metric x increases to the right of the aligned master image, while metric y reverses the image's downward axis.

Processing Commands

First transform the pixel centers to meters:

python coordinate_transform.py --input RAW_DETECTION_DATA.csv --calibration SCENE_CALIBRATION.json --source SOURCE_VIDEO_KEY --output METRIC_TRAJECTORIES.csv

Add verified lane and direction annotations to that output, then reconstruct the trajectories:

python reconstruct_trajectories.py --input METRIC_TRAJECTORIES_WITH_LANES.csv --fps SOURCE_FPS --source SOURCE_VIDEO_KEY --output RECONSTRUCTED_TRAJECTORIES.csv --rejected REJECTED_TRAJECTORIES.csv --audit SCREENING_SUMMARY.csv

Estimate lengths from the original six-column detections using the same calibration, and merge them into the retained trajectories:

python estimate_vehicle_length.py --input RAW_DETECTION_DATA.csv --calibration SCENE_CALIBRATION.json --source SOURCE_VIDEO_KEY --output VEHICLE_LENGTH_ESTIMATES.csv
python merge_vehicle_lengths.py --input RECONSTRUCTED_TRAJECTORIES.csv --lengths VEHICLE_LENGTH_ESTIMATES.csv --output BASIC_TRAJECTORY_DATA.csv

Generate the two semantic layers independently:

python annotate_risk.py --input BASIC_TRAJECTORY_DATA.csv --output RISK_SEMANTIC_DATA.csv
python annotate_interactions.py --input BASIC_TRAJECTORY_DATA.csv --pairs CANDIDATE_PAIRS.csv --parameters PR_PARAMETERS.csv --output INTERACTION_SEMANTIC_DATA.csv

Each constant-frame-rate recording should be processed separately. When multiple recordings are combined, source identity, vehicle identifiers, and physical time should be retained to ensure consistent mapping across processing stages.

Screening and Reconstruction Settings

The released reconstruction code applies trajectory-quality screening based on observed duration, observation count, spatial extent, and average path speed. The basic screening criteria retain tracks with an observed duration N/FPS of at least 0.5 s and at least 15 valid observations. Low-motion trajectories are additionally evaluated using duration, coordinate extent, and average path speed. Here, N counts observations before interpolation, and average path speed is calculated as the sum of consecutive observed travel distances divided by N/FPS.

These rules provide a consistent quality-screening procedure for short-duration and low-motion trajectories. The resulting screening outputs and per-vehicle summary support transparent review of the retained trajectory set before reconstruction.

For retained tracks, exactly two valid coordinate observations select linear interpolation and three or more select PCHIP. Lane and direction values use forward fill followed by backward fill; vehicle identity remains constant within each trajectory. Position smoothing uses sEMA with a default time parameter of 0.6 s, followed by positional differences for velocity and acceleration.

Newly reconstructed outputs additionally retain Source_ID, Original_Vehicle_ID, and Original_Frame, together with Time_s. Their Delta_t_s is defined from the second trajectory record onward as 1/FPS. These lineage and timing fields support source tracing and temporal reconstruction during customized reprocessing.

Vehicle-Length Estimation Settings

The estimator constructs bounding-box corners internally from the center coordinates and half the width/height. It transforms all four corners to the road-aligned Frenet-based metric coordinate system and uses the longitudinal maximum-minus-minimum extent as a frame-level length candidate, providing a direct geometric length estimate from the supplied bounding boxes.

Quality-control screening removes nonfinite measurements and nonpositive box dimensions before aggregation. With at least four valid candidates, a 1.5-IQR filter is applied, and the 25th percentile of the retained candidates is assigned as the fixed vehicle length when at least five observations are retained. The output lateral extent provides an additional bounding-box diagnostic. The multi-frame aggregation and IQR-based screening improve robustness to frame-level bounding-box variation, while the six pixel-space fields are combined with metric calibration to estimate physical vehicle length.

Semantic-Annotation Inputs and Interpretation

Risk annotation uses physical time, lane/direction information, and a valid fixed length for each vehicle. The current script uses the nearest same-lane leader for TTC/DRAC and trajectory crossings associated with cut-in events for PET. PET uses future observations and is an event-based retrospective measure. The default DRAC cap is 15 m/s². An optional vehicle-length fallback setting is available when required, and TTC and PET values are interpreted according to their metric-specific applicability conditions.

Candidate-pair input requires Frame, EV_Vehicle_ID, POV_Vehicle_ID, x_local, y_local, v_rel, cangle. Pairs must refer to valid vehicles in the same source/frame, travel in the same direction, lie within 50 m, and have positive scalar relative speed. The interaction angle described in the manuscript is a continuous quantity. For the processing implementation, angles from 5° to 175°, inclusive, are encoded as cangle = 1, and angles outside this interval as cangle = 0. This conversion must be completed before supplying the candidate-pair file to the annotation script. The value 1 selects Lane Changing/Weaving; the value 0 selects Car Following for vehicles in the same lane and Parallel Driving otherwise.

PR parameter input requires cangle, round_v, rx_plus, rx_minus, ry_plus, ry_minus, bx_plus, bx_minus, by_plus, by_minus. Scale and shape parameters must be positive. The script rounds relative speed to one decimal place and selects the nearest parameter bin within the matching cangle group. Candidate-pair inputs and PR parameter files must use the same angle encoding, local-coordinate convention, and calibration procedure.

All supplied candidate pairs receive annotations. Interaction Strength > exp(-1) corresponds to Is Interacting = 1, while candidate pairs below this threshold remain available with their corresponding interaction annotations. Records in the default no-candidate state are encoded with Interaction_Target_ID = -1, interaction strength 0, interaction flag 0, and interaction type Safe.

Processing Inputs and Reprocessing

The released package provides executable implementations of the principal post-processing stages, including coordinate transformation, trajectory screening and reconstruction, kinematic calculation, vehicle-length estimation, risk annotation, and interaction annotation. These modules use the corresponding scene-specific calibration information, lane and direction information, candidate-pair data, and PR parameters described in the input specifications above. The supplied templates, input schemas, and processing instructions provide a structured workflow for applying these modules to the released data or to user-prepared inputs.

Users can analyze the published processed data directly or use the released raw records and processing modules for customized reprocessing. The required scene/source inputs, parameter settings, and processing procedures are described above and in the processing README.

Reading and Selecting Data

Download the CSV for the observation group and processing layer needed for your analysis. The interaction files may contain multiple candidate-pair rows for a given ego-vehicle/frame combination; ego/frame and target identifiers can therefore be used jointly in pair-level analyses.

For example, after downloading a basic trajectory file:

import pandas as pd

filename = "Basic_Trajectory_Data/Weaving Segment_3.csv"
for chunk in pd.read_csv(filename, chunksize=100_000):
    print(chunk.head())
    break

Use file/source identity together with vehicle and frame identifiers when linking data. For interaction rows, also retain Interaction_Target_ID; unique ego/frame records can be used for trajectory counting and time-axis reconstruction. The geometry images provide contextual documentation of roadway layout and channelization.

Users may define train/validation/test splits by recording, vehicle, or scenario according to the specific predictive-modeling task, with separation maintained between neighboring frames and overlapping vehicle trajectories. Files from different processing layers have different columns and can be loaded according to their corresponding schemas.

Relationship Between Data Components

The repository provides multiple data levels so that users can select the level most appropriate for their research objective.

Raw_Data provides low-level frame-wise image-coordinate and bounding-box information and is intended for users who wish to independently perform trajectory post-processing.

Basic_Trajectory_Data provides reconstructed physical vehicle trajectories and is suitable for vehicle kinematics, lane-changing behavior, traffic-flow modeling, trajectory prediction, and simulation calibration.

Risk_Semantic_Data retains the basic trajectory information and supplements it with TTC, PET, DRAC, selected risk-target information, and explicit risk labels.

Interaction_Semantic_Data retains the basic trajectory information and supplements it with vehicle-pair relative states, PR-based interaction strength, effective-interaction labels, and interaction-type annotations.

Weaving_Segment_Geometry_Images provides the roadway and channelization context required to interpret lane, direction, and merge/diverge-related variables.

The processed risk and interaction files already retain the corresponding basic trajectory variables, allowing users to directly select the data component matching their research objective.

Technical Validation

The accompanying manuscript reports evaluation of U-EASWS from multiple complementary perspectives.

Vehicle detection and multi-object tracking performance were examined under dense UAV-based traffic observations.

Trajectory continuity and spatial consistency were evaluated after coordinate registration, reconstruction, and smoothing.

Kinematic consistency was assessed using reconstructed position, velocity, and acceleration profiles, with particular attention to trajectory-boundary regions where derivative variables are most sensitive to positional variation. The validation results provide complementary evidence for the continuity and physical consistency of the reconstructed trajectories.

The spatial distributions of lane-changing events were further examined across all seven Type-A short weaving segments. The results show systematic relationships between lane-changing behavior and roadway geometry, lane position, travel direction, and channelization constraints across the observation areas.

The risk semantic layer was evaluated through the spatial and temporal characteristics of TTC, PET, and DRAC.

The interaction semantic layer was evaluated by examining the relationship between PR-based interaction strength and microscopic vehicle responses, as well as the spatial distribution of effective interactions across the weaving areas.

Usage Notes

Users interested primarily in vehicle trajectories and kinematic analysis can directly use Basic_Trajectory_Data.

Users studying traffic conflicts and surrogate safety should use Risk_Semantic_Data.

Users studying vehicle-pair interaction, spatial interaction strength, or interaction type should use Interaction_Semantic_Data.

Users who wish to perform independent trajectory reconstruction or alternative processing pipelines should begin with Raw_Data.

For files containing bidirectional traffic, users should first distinguish traffic streams using Direction before conducting lane-level, vehicle-level, or weaving-segment-specific analyses.

The fields Lane number, Dist_to_Merge_m, and Dist_to_Diverge_m should be interpreted together with the corresponding channelization image in Weaving_Segment_Geometry_Images.

TTC and DRAC represent frame-wise conflict-related states, whereas PET is an event-level measure associated with identified cut-in events.

PR-based Interaction Strength characterizes proximity-induced interaction, while the risk semantic fields describe conflict-related states. The two semantic layers therefore provide complementary information for microscopic traffic analysis.

Comparative analyses across WS1–WS7 can incorporate roadway geometry, channelization configuration, lane number, traffic direction, observation duration, weaving length, and traffic conditions. Scenario-level analyses can also be conducted before cross-scenario aggregation or normalization to preserve the characteristics of each weaving segment.

Data Interpretation and Usage Considerations

The released Raw_Data provides frame-level vehicle detection and tracking outputs representing an earlier processing stage than the reconstructed trajectory data.

These records provide image-space vehicle positions and detection-box dimensions that support independent trajectory post-processing and customized methodological workflows.

The road-aligned Frenet-based coordinates are scene-specific metric coordinates designed to describe longitudinal and lateral vehicle motion within each weaving scenario.

Vehicle_Length_m is an estimated physical vehicle attribute derived from multi-frame detection bounding boxes transformed into the road-aligned Frenet-based metric coordinate system.

The interaction-strength annotation quantifies proximity-induced interaction and complements the TTC-, PET-, and DRAC-based risk semantic information.

The dataset covers seven real-world Type-A short weaving scenarios in Changchun, China, providing a focused empirical basis for microscopic traffic-behavior, safety, and interaction research and supporting comparative analyses across different roadway geometries and channelization configurations.

Licensing Information

U-EASWS is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

Under this license, users may share and adapt the dataset material for any purpose, provided that appropriate credit is given to the original dataset creators and any changes are indicated.

Dataset Access

The released data and principal code are available in the U-EASWS repository.

The repository currently contains:

• Weaving_Segment_Geometry_Images
• Raw_Data
• Basic_Trajectory_Data
• Risk_Semantic_Data
• Interaction_Semantic_Data
• Principal Processing Code