SIND: A Drone Dataset at Signalized Intersection in China

I. INTRODUCTION

• Intersections combine complex topology and dynamic traffic, while motion prediction needs realistic trajectories and context such as road rules and traffic-light states.The paper notes that signal states directly affect traffic-participant motion and are important to integrate into prediction models.Source: §I · p. 2
• SIND records a signalized Tianjin intersection with traffic-light states, an HD map, diverse traffic participants, behavior labels, and traffic-light violations.The paper presents the dataset as relevant to mixed-traffic prediction, behavior modeling, scenario construction, and related research.Source: §I · p. 2

II. RELATED WORK

• NGSIM's Lankershim Boulevard data cover 30 minutes of vehicle trajectories but omit vulnerable road users, unlike SIND.The passage says NGSIM is mainly oriented to traffic simulation.Source: §II · p. 2
• INTERACTION spans 16.5 hours and diverse road structures, but provides few vulnerable road users, limited motion features, and no signal-state information [15].Its reported vulnerable-road-user share is about 5%, and the dataset provides velocity in global coordinates as an example of its motion features.Source: §II · p. 2
• The inD dataset has 13,599 trajectories over 10 hours at unsignalized intersections, while rounD and openDD cover unsignalized roundabouts,, [33].The passage reports vulnerable-road-user shares below 5% for rounD and openDD.Source: §II · p. 2

• The pipeline includes selecting a recording location.Location selection is one of the listed construction stages.Source: §III · p. 3
• The pipeline includes recording data at the selected site.Data recording is one of the listed construction stages.Source: §III · p. 3
• The pipeline also covers data processing and map generation.Both are listed as construction stages.Source: §III · p. 3

A. Selection of recording location

• The selected intersection uses two-phase lights, with left-turning and straight-going vehicles sharing a green phase that leads to frequent interactions and conflicts.The site was chosen to capture typical traffic patterns as well as non-compliant behavior, including violations by vulnerable road users.Source: A. Selection of recording location · p. 3

B. Data recording and preprocessing

• A DJI Mavic 2 records intersection video at 29.97 Hz, while a sidewalk camera records traffic-light states unavailable from the drone view.The drone carries a 4K (3840x2160) camera.Source: B. Data recording and preprocessing · p. 3
• The 23 recordings span 15 days, cover morning and afternoon periods and varied weather, and last 10–22 minutes each.The drone hovers at 75 m, a height chosen for view coverage and image clarity.Source: B. Data recording and preprocessing · p. 3
• A synchronization frame in each video pair supports alignment of the traffic-light and drone-camera timelines to a unified timestamp.The timestamps are aligned using video-editing software.Source: B. Data recording and preprocessing · p. 3
• Downsampling keeps one-third of frames, and SURF-based matching registers frames to the first frame and HD map to reduce drone-motion disturbance.The method addresses slow participant movement between frames in the original video.Source: B. Data recording and preprocessing · p. 3

• Detection must obtain each traffic participant's class and position in every frame to create trajectories.Source: C. Detection and correction · p. 3
• Pedestrians are annotated as points and vehicles as axis-oriented rotated rectangles, using YOLOv5-based detection with manual correction for inaccurate cases below 5%.Vehicle size and orientation are treated as important features.Source: C. Detection and correction · p. 4
• A method from corrects relief displacement because elevated drone imaging is not orthographic, an effect especially evident for trucks and buses.The passage attributes outward image offset from the imaging center to object height.Source: C. Detection and correction · p. 4

D. Tracking and Postprocessing

• IOU matching and linear Kalman filtering track participants; predicted positions fill gaps from occlusions, with manual merging for a few duplicate tracks.The authors note tree shade as one source of occlusion.Source: D. Tracking and Postprocessing · p. 4
• The HD map is drawn in Lanelet2 format from a reference image, with traffic participants and map elements converted to the ground coordinate system.Source: D. Tracking and Postprocessing · p. 4
• Tracks are converted to ground coordinates using a similar transformation, with scale measured from ground references and translation and rotation obtained by registering first frames.The conversion assumes drone-gimbal stabilization makes camera pitch negligible.Source: D. Tracking and Postprocessing · p. 4
• RTS smoothing applies kinematic constraints using a bicycle model for vehicles and a point-mass model for pedestrians, refining trajectories and estimating velocity and acceleration.The passage states that smoothing reduces the influence of detection error.Source: D. Tracking and Postprocessing · p. 4

E. Datasets and Formats1

• SIND records participant position, velocity, and acceleration, with vehicle heading and yaw also provided in the ego-vehicle coordinate system.Source: E. Datasets and Formats1 · p. 4
• The dataset includes states for eight traffic lights, encoded as 0 for red, 1 for green, and 3 for yellow.Source: E. Datasets and Formats1 · p. 4
• A Lanelet2 map in OSM format and an aerial image describe static elements of the intersection.Fig. 4 shows the Lanelet2-format map.Source: E. Datasets and Formats1 · p. 4, Figure 4 · p. 4
• Record metadata includes participant start and end frames, vehicle crossing types, traffic-light violations, collection time, weather, and participant counts.Source: E. Datasets and Formats1 · p. 4

A. Scale and number of categories

• 13,248 trajectories record more than 99.8% of traffic participants at the intersection, excluding a very small number of special objects such as pets and carts.Source: A. Scale and number of categories · p. 4
• 420 minutes of data comprise 23 records, each lasting 8–20 minutes, recorded between July 28 and August 11, 2021.Recording conditions included sunny, cloudy, and post-rain weather.Source: A. Scale and number of categories · p. 4
• The seven classes are cars, trucks, buses, tricycles, bikes, motorcycles, and pedestrians; motorcycles, bicycles, tricycles, and pedestrians are identified as VRUs.Fig. 5 depicts the number of participants in each class.Source: A. Scale and number of categories · p. 4, Figure 5 · p. 5

B.Traffic light violations

• A red-light violation requires the vehicle to cross the stop line and continue to the opposite side after the red light comes on; a yellow-light violation occurs when it crosses after yellow is already on.Over-the-line parking without driving to the opposite side is excluded from the red-light definition.Source: B.Traffic light violations · p. 5
• Violations are identified using HD-map behavior classification, signal states, and the frame when a vehicle crosses the stop line; non-motor vehicles use the zebra-crossing border instead.Only left-turning and straight-going motor vehicles are considered because right turns are allowed at all times in China.Source: B.Traffic light violations · p. 5
• Fig. 6 reports category counts as car (29/122), bus (0/0), truck (3/0), bicycle (698/69), motorcycle (1072/86), and tricycle (108/14).Red denotes running a red light, yellow denotes running a yellow light, and green denotes other behavior.Source: Figure 6 · p. 5

C. Comparison and Discussion

• INTERACTION has the largest scale, with 18,471 intersection trajectories, but provides no signal-state information; inD was collected at four unsignalized intersections.The passage also questions INTERACTION's completeness for pedestrians and bicycles based on a provided snapshot and corresponding files.Source: C. Comparison and Discussion · p. 5
• SIND provides comprehensive size and motion-state features for motor and non-motor vehicles, and its scale is close to inD while its traffic activities are more intensive.The authors describe SIND as having the most complete motion parameters among the compared works.Source: C. Comparison and Discussion · p. 6
• SIND distinguishes vehicle heading from yaw, defining heading by centroid velocity direction and yaw by the vehicle's longitudinal-axis direction.The paper relates the angles by θ = β + ψ, where β is the center-of-mass sideslip angle.Source: C. Comparison and Discussion · p. 6
• About 62.6% of SIND trajectories are from vulnerable road users, and visual review found quality close to or possibly better than inD, especially for small bicycles and motorcycles.The authors note that the quality comparison is based on visual observation of demonstration videos.Source: C. Comparison and Discussion · p. 6
• At this two-phase intersection, left-turning and straight-going vehicles share a signal phase, creating interactions and conflicts relevant to autonomous-driving research.Left-turning vehicles have no right of way over straight-driving vehicles in this setting.Source: C. Comparison and Discussion · p. 6
• SIND records many non-motor-vehicle traffic-light violations, whose proportion far exceeds motor-vehicle violations; violations are associated with intersection size and traffic flow.The authors report illegal crossings as more likely at smaller intersections and in sparse traffic.Source: C. Comparison and Discussion · p. 6

V. CONCLUSION

• SIND provides comprehensive traffic-participant motion features, traffic-light states, and HD maps from a signalized intersection in Tianjin, China.The dataset also labels vehicle traffic-light violations according to traffic rules in China.Source: §V · pp. 6–7
• SIND includes a high proportion of vulnerable road users and presents traffic-light violations by non-motor vehicles.The authors identify these features as distinguishing SIND from other works.Source: §V · p. 7
• The authors expect SIND to support research on autonomous driving, especially motion prediction in mixed traffic and behavioral modeling of vulnerable road users.Source: §V · p. 7

VI. LIMITATION AND FUTURE WORK

• Because SIND was collected at a single location, it is not conducive to analyzing how location and road structure affect traffic-participant behavior.Planned future collection spans intersections in different Chinese cities and considers regional characteristics and intersection types.Source: §VI · p. 7
• Pedestrian behavior-type and violation-type labels are not provided because pedestrian trajectories are difficult to generalize into a few types.The authors identify pedestrian behavior generalization and labeling as follow-up work.Source: §VI · p. 7
• Future work also aims to advance urban autonomous-driving scenario tests, VRU behavioral analysis, and prediction and decision-making research in comprehensive scenarios.The authors say these results will appear in subsequent publications.Source: §VI · p. 7