A Systematic Evaluation of Infrastructure-Based Radar System for Highway Traffic Monitoring

I. INTRODUCTION

• Fine-grained vehicle positions and dynamics support traffic-flow and driving-behavior research as well as real-time state estimation, signal optimization, and safety management.Source: §I · p. 1
• Infrastructure sensors are designed to monitor all vehicles in their detection zones, whereas probe-vehicle trajectories can be sparse and partial because of limited market penetration.Source: §I · p. 1
• A systematic, large-scale assessment of radar detection across multiple levels remains unexplored, while prior studies report aggregate counts or validate only selected vehicles.Radar can offer longer range and greater robustness, but coarse angular resolution, multipath reflections, and reduced sensitivity to stationary or slow-moving objects pose monitoring challenges.Source: §I · p. 1
• DRaT pairs naturalistic vehicle trajectories from radar and drone sensors to systematically assess radar detection and tracking against drone-derived ground truth.The dataset was collected at a highway merging area in Fort Worth, Texas, and evaluates individual detection, trajectory tracking, and macroscopic traffic parameters.Source: §I · pp. 1–2

A. Study area

• The study covered about 140 m of eastbound roadway with three regular lanes and one merging lane, during 43 minutes of free flow and 37 minutes of congestion.Free-flow and congested periods differed in average density, space-mean speed, and traffic volume.Source: A. Study area · p. 2

1) Radar:

• The commercial radar uses a 24.25–24.50 GHz K-band FMCW architecture and provides long detection range and a wide horizontal field of view.Its large antenna aperture provides angular resolution to support vehicle-dimension estimation.Source: 1) Radar · p. 2

2) Drone:

• The drone supports multi-constellation GNSS positioning and stable hovering with sub-meter accuracy, and records high-resolution video using a 3-axis stabilized gimbal.Source: 2) Drone · p. 2
• The drone hovered at 122 m and recorded 4K video at 30 fps; its footage was aligned to include the radar location to maximize overlapping sensor coverage.The drone covered about 140 m longitudinally, and radar detections were restricted to the same range for consistent comparison.Source: 2) Drone · pp. 2–3
• The radar was mounted about 5.5 m high with a 15° downward tilt toward traffic, balancing detection range, occlusions, and near-field blind regions.Fig. 1 depicts the study area, sensor configuration, and radar mounting on a portable power generator.Source: 2) Drone · p. 2, Figure 1 · p. 2

A. Drone

• The drone videos were stabilized frame by frame, then registered to a georeferenced orthophoto to transform image coordinates into real-world coordinates.The orthophoto used a ground sampling distance of about 0.059 m/pixel and was georeferenced with ground control points.Source: A. Drone · p. 3
• Using the orthophoto as the spatial reference makes trajectories largely independent of drone GPS errors and hovering inaccuracies.Ground-truth accuracy is mainly determined by video stabilization, orthophoto registration, and image-to-world transformation quality.Source: A. Drone · p. 3
• Drone trajectories were refined with per-frame greedy NMS and downsampled from 30 Hz to the radar’s exact 8 Hz timestamps for one-to-one temporal correspondence.The NMS procedure suppresses later boxes within 1.5 m or with IoU ≥0.1, or boxes at least 10% engulfed by a kept box.Source: A. Drone · p. 3
• After radar-timestamp alignment, the drone footage yielded 392,707 individual vehicle detections; Fig. 2 shows examples, and high-fidelity detection and robust georeferencing support their use as ground truth for radar evaluation.Source: A. Drone · p. 3

B. Radar

• Radar outputs vehicle positions, speeds, dimensions, and headings at 8 Hz.The built-in algorithm also provides object tracking.Source: B. Radar · p. 3
• Radar detections and drone data undergo the same non-maximum suppression procedure to remove redundant false positives.Source: B. Radar · p. 3
• Radar positions are converted to GPS coordinates using the installation location and azimuth, with RTK-corrected GPS used to improve georeferencing.Source: B. Radar · p. 3
• Post-processing yielded 5,430 vehicle trajectories comprising 282,712 individual detections.Source: B. Radar · p. 3

A. Individual vehicle detection

• 78.3% precision and 56.8% recall summarize overall radar detection performance at an IoU threshold of 0.1.The table summarizes performance under different traffic conditions.Source: A. Individual vehicle detection · p. 4, Table III · p. 4
• Congestion degrades recall through reduced sensitivity to slow targets, vehicle occlusion, and restricted target separation, while multipath reflections and fragmented truck detections add false positives.Fig. 4 shows semi-truck occlusions; Fig. 5 illustrates large vehicles split into multiple smaller detections.Source: A. Individual vehicle detection · p. 4, Figure 4 · p. 4, Figure 5 · p. 4
• Beyond 60 m, precision drops from 0.82 to 0.66 and recall from 0.73 to 0.33; within 30 m, recall is 0.55.Fig. 6 presents precision and recall by longitudinal segment; near-field performance is linked to the radar’s 15° downward boresight, while longer-range errors combine weaker signal, coarser crossrange resolution, localization error, and occlusion.Source: A. Individual vehicle detection · p. 4, Figure 6 · p. 4
• Spatially, recall decreases toward outer lanes, farther distances, and the near-field region, while precision stays relatively stable except near the boundaries.Fig. 7 shows the spatial distribution with the radar at the origin, oriented along the positive vertical axis.Source: A. Individual vehicle detection · p. 4, Figure 7 · p. 4

B. Trajectory tracking

• Radar-derived trajectories tend to be shorter than drone-derived trajectories, likely because of more missed detections at longer distances.Fig. 8 compares the duration distributions.Source: B. Trajectory tracking · p. 5
• MOTA is 0.535 and IDF1 is 0.699 overall; 47% of ground-truth trajectories are mostly tracked, while 38% are partially tracked and 14% mostly lost.Identity precision is 0.83 and identity recall is 0.60; the table reports tracking performance under different traffic conditions.Source: B. Trajectory tracking · p. 5, Table IV · p. 5
• Under congestion, MOTA decreases from 0.601 to 0.509 and IDF1 from 0.766 to 0.672, while identity switches increase from 309 to 789.The proportion of mostly tracked trajectories falls from 55% to 38%; MOTP remains 1.79 m across traffic conditions.Source: B. Trajectory tracking · p. 5

C. Macroscopic traffic parameter estimation

• Macroscopic parameters are estimated using 30-second windows, and Fig. 9 visualizes radar estimates over a 15-minute period.Source: C. Macroscopic traffic parameter… · p. 5
• Space-mean speed estimation has an average MAE of 1.37 km/h and MAPE below 4%, while density and volume MAPE are approximately 23%.Source: C. Macroscopic traffic parameter… · p. 5
• Missed detections are the primary source of density and volume error, producing systematic underestimation; errors are fairly consistent over time rather than dominated by occasional large deviations.The latter is indicated by the relatively small difference between RMSE and MAE.Source: C. Macroscopic traffic parameter… · p. 5

• DRaT provides drone-derived ground-truth trajectories for systematic radar evaluation and can support reproducible work on identifying, characterizing, and potentially mitigating sensing errors.The dataset is particularly relevant to merging areas with lane-changing interactions and dense vehicle maneuvers.Source: §V · pp. 5–6
• Detection degrades in congestion and at longer sensing distances, while tracking also deteriorates under congestion; radar is suited to motion-centric, near-to-mid-range monitoring but warrants caution for standalone complete detection and accurate counting.The discussion links tracking deterioration under congestion to increased occlusion and reduced target separation.Source: §V · p. 6
• Radar–vision fusion can weight radar more heavily for speed-related states and vision for vehicle class and dimensions, especially when dense traffic limits radar resolution.The discussion associates radar with reliable velocity measurements and robust perception under challenging weather and lighting conditions.Source: §V · p. 6
• Evaluation is constrained to a commercial radar with a built-in algorithm, one mounting height, and one highway segment; other road types remain to be assessed.The authors name curves, ramps, and complex interchanges as settings for further evaluation.Source: §V · p. 6