"Event-Based Camera Infrastructure Node Dataset"

"SummaryThis dataset is event-based camera data for moving object detection from a static, surveillance-style setup. It targets the use of event-based vision in ITS infrastructure nodes (e.g. cooperative collision avoidance). It contains three object classes (pedestrian, cyclist, car) and three sensor sensitivity levels (RVD: 100 mV, 75 mV, 56 mV), with 2D bounding box labels. It was used to benchmark clustering-based detectors and to study the effect of RVD and accumulation time.PurposeTo evaluate whether an event-based camera with classical (clustering) detection can support ITS infrastructure node applications (e.g. low-latency moving object detection). No existing public dataset at the time offered multi-class surveillance event data with multiple RVDs; this dataset fills that gap for algorithm comparison and parameter studies (RVD, accumulation time, clustering methods).SensorCamera: PROPHESEE ONBOARD event-based camera.Resolution: 640\u00d7480 pixels (15 \u00b5m pixel pitch).Temporal resolution: up to ~10 kHz equivalent.Dynamic range: >120 dB.Mounting: 4 m height, 15\u00b0 downward tilt; 103\u00b0 diagonal FOV lens (surveillance-style).Scenario and collectionSetup: Controlled area; only the target object moves along a fixed path per run (no unrelated traffic).Classes: Pedestrian, cyclist, car. Each class follows the same path three times, giving nine sequences per RVD.RVD (sensitivity): Three RVDs: 100 mV (low noise, ~13 Keps), 75 mV (medium, ~111 Keps), 56 mV (high noise, ~19 Meps). Nine sequences per RVD (27 sequences in total).Sampling for labels: Events are accumulated and labelled at 5 Hz to maximise inter-frame difference for detection (no tracking).AnnotationsAccumulated event frames (from raw events) were manually annotated in CVAT.A co-located frame-based RGB camera was used as reference.Labels: 2D bounding boxes (frame index, x, y, width, height).Minimum box size: 20\u00d720 px.Shadows are included inside the corresponding object box.Split: First run per class = tuning; second and third runs = test.Dataset sizeDuration: ~23 minutes total (~1,402 s).Labelled frames: 7,009.By noise level: Low 456 s \/ 2,280 frames; medium 469 s \/ 2,344 frames; high 477 s \/ 2,385 frames.By class: Pedestrian 581 s \/ 2,905 frames; cyclist 411 s \/ 2,055 frames; car 410 s \/ 2,049 frames.FormatsRaw events: CSV with (Timestamp, X, Y, Polarity).Ground truth: CSV with (Frame Number, X, Y, Width, Height)."