UHCTD

The University of Houston Camera Tampering Dataset (UHCTD) is a large-scale resource designed to train and evaluate algorithms for detecting unauthorized or accidental changes in surveillance views. Tampering—whether caused by natural phenomena like sunlight reflection and fog, or malicious human intent like spray painting and lens blocking—compromises the integrity of public safety systems.

Dataset Overview

UHCTD features over 26GB of data captured from two high-resolution outdoor cameras (2048x1536 and 1280x960). We utilized advanced image processing techniques to synthesize three primary categories of tampering into real-world surveillance footage:

• Covered: View blocked by objects (cardboard, hands) or natural accumulation (dust, webs).
• Defocussed: Blurred views due to lens fogging or intentional focus manipulation.
• Moved: Changes in the camera’s viewpoint caused by strong winds or manual redirection.

Benchmarking

The project includes a comprehensive evaluation of standard deep learning architectures for tampering detection, including:

• AlexNet
• ResNet-18 / ResNet-50
• DenseNet-161

Resources & Development Kits