Anomaly Detection Dataset

Home » Dataset Download » Anomaly Detection Dataset

Anomaly Detection Dataset

Datasets

File

Anomaly Detection Dataset

Use Case

Anomaly Detection Dataset

Description

Explore the UCF-Crime Anomaly Detection Dataset, featuring 128 hours of real-world surveillance footage with 13 high-impact anomalies. Ideal for training AI models in public safety and security applications.

Description:

The UCF-Crime Dataset is one of the largest publicly available datasets designed for anomaly detection in video surveillance systems. It contains an extensive collection of 128 hours of video footage, captured from real-world surveillance cameras, offering a robust and diverse dataset for training AI models in detecting and recognizing abnormal activities in public spaces.

This dataset comprises a total of 1,900 long and untrimmed videos, reflecting the unfiltered and realistic nature of security surveillance. The videos capture 13 distinct types of anomalous activities that pose significant threats to public safety. These activities include Abuse, Arrest, Arson, Assault, Road Accidents, Burglary, Explosion, Fighting, Robbery, Shooting, Stealing, Shoplifting, and Vandalism. These anomalies were selected based on their real-world impact, ensuring that the dataset remains highly relevant for developing security and public safety applications.

Download Dataset

The UCF-Crime dataset is particularly suited for two primary tasks:

General Anomaly Detection: In this task, all 13 types of anomalies are grouped together as abnormal events, while all other activities are classified as normal. This is ideal for training models to detect the presence of any irregular or suspicious behavior, regardless of its specific nature.
Anomalous Activity Recognition: This task is more granular, with models being trained to identify and classify each type of anomalous activity separately. This allows for the development of more sophisticated systems capable of recognizing specific threats, such as detecting a robbery versus identifying a road accident.

The videos in this dataset are long and untrimmed, closely mirroring the real-world scenarios in which anomalies occur amidst large volumes of mundane activities. This makes it ideal for training models that need to not only detect abnormal behavior but also sift through hours of normal footage without triggering false positives.

Key Features of the UCF-Crime Dataset:

Large-Scale Data: 1,900 videos amounting to 128 hours of footage, providing ample data for training deep learning models in anomaly detection.
Diverse Anomalies: Includes 13 realistic and high-impact anomalies, offering a wide range of events that are crucial for public safety monitoring.
Real-World Surveillance Footage: Untrimmed and lengthy videos provide a realistic environment for model training, improving robustness and reducing false alarms.
Two Levels of Task Complexity: Supports both general anomaly detection and specific activity recognition, allowing for flexible use in various AI applications.

Applications:

This dataset is designed for use in the following areas:

Public Safety: Developing AI systems to enhance safety and security by detecting dangerous situations in public spaces, including parks, streets, and transportation hubs.
Surveillance System Enhancement: Training video surveillance systems to automatically detect and alert authorities to anomalous activities, improving response times in emergencies.
Smart Cities: Integrating anomaly detection into the infrastructure of smart cities to ensure the safety of citizens in real-time.
Law Enforcement: Assisting police and security agencies in identifying and responding to criminal activities such as burglary, assault, and vandalism.