Ensuring safety in urban transportation requires robust anomaly detection, yet existing datasets often lack diverse, real-world operational data from public transit environments. This limitation makes it challenging for models to generalize across varying lighting conditions, occlusions, crowd densities, and vehicle interiors, which are critical factors in real-world anomaly detection. To address this gap, we introduce the Traffic Anomaly Detection Dataset (TADD), a large-scale, multimodal dataset designed for safety-critical anomaly detection in urban transit settings. The "The TADD dataset includes RGB surveillance footage from bus platforms and RGBD sensor data from taxi cabins, covering 29 fine-grained anomaly categories across 212 video sequences and over 20,000 manually annotated frames. Additionally, the dataset contains 17 anomaly categories with unlabeled RGBD sensor data from bus cabins, 9 of which represent new categories not found in taxi cabins. In total, the dataset features 38 distinct anomaly categories. The labeled footage spans 5.67 hours, while the unlabeled footage from bus cabins alone contributes 5.23 hours. By incorporating scene, pose, and occlusion variations, TADD provides a challenging testbed for evaluating anomaly detection models. We establish benchmark evaluations with state-of-the-art detection and classification frameworks, highlighting performance gaps in temporal reasoning, spatial context understanding, and multimodal fusion. These findings suggest that further improvements are needed to enhance model robustness in real-world transit applications. By offering a diverse and operationally realistic dataset, TADD contributes to the advancement of intelligent transportation safety monitoring.
Representative scenarios in the TADD dataset. Different colored borders represent distinct camera viewpoints, with 12 total viewpoints: 10 at the public transit platforms, 1 in the taxi cabin and 1 in the bus cabin. The dataset includes multiple categories of anomalous interactions, encompassing individual passenger behaviors, passenger-to-passenger interactions, and passenger-to-driver interactions, as well as categories for anomalous objects.
TADD improves upon existing anomaly detection datasets by offering a broader range of interaction types and a unique combination of platform and in-vehicle scenarios.
Video duration (in seconds) for various TADD labels. Note: The TADD-Cabin data includes the total seconds for both RGB and RGBD video types