"TrafficNet-2024"

"This dataset presents a large-scale benchmark for traffic incident detection, constructed through comprehensive feature extraction and processing of 2024 traffic data from the California Performance Measurement System (PeMS). The dataset encompasses traffic flow data collected from 841 detectors across 20 freeways in California, capturing the period from January through October 2024. It includes 38,250 traffic incidents extracted from original incident records, which have been re-categorized into 118 distinct anomaly families by expanding the original seven-class taxonomy.A key characteristic of this dataset is its severe class imbalance, with an imbalance ratio (IR) of 5201.87, making it particularly suitable for evaluating imbalanced learning algorithms. To ensure robust evaluation of few-shot cross-class generalization, the dataset employs a family-wise splitting strategy where anomaly families are shuffled and partitioned into training (60%), validation (20%), and test (20%) subsets using a random seed of 42. This dataset is designed to support research in few-shot learning, imbalanced classification, and meta-learning approaches for traffic incident detection."