Multimodal Video Analysis for Crowd Anomaly Detection Using Open Access Tourism Cameras

Dataset Description This repository contains a dataset of object detection and instance segmentation events recorded over a 34-day period, from September 19, 2023 to October 23, 2023. The dataset comprises 50,384 individual records, capturing temporal, spatial, and classification data for objects identified in a open access video broadcast from the town of Morella, Spain. Dataset Structure The repository contains a single file, detections.csv. detections.csv serves as the primary dataset file. It encapsulates granular detection events, where each record includes the temporal timestamp, categorical class ID, model confidence metric, and spatial geometry. Comprehensive details regarding these columns are provided in the section below. Dataset Codebook The detections.csv file contains 8 columns. Column Name Data Type Description time String Timestamp of the detection in YYYY-MM-DD HH:MM:SSformat. class_id Integer Integer representing the object category based on the COCO dataset. Detected classes include: Class 0 (person), Class 1 (bicycle), Class 2 (car), Class 3 (motorcycle), Class 5 (bus), and Class 7 (truck). confidence Float Model certainty score (0.0 - 1.0). xmin Float X-coordinate of the bounding box's top-left corner (pixels). ymin Float Y-coordinate of the bounding box's top-left corner (pixels). xmax Float X-coordinate of the bounding box's bottom-right corner (pixels). ymax Float Y-coordinate of the bounding box's bottom-right corner (pixels). segment_mask String Contains a stringified list of [x, y] pixel coordinates that form a closed polygon around the detected object.