Human Activity Recognition (HAR) has become increasingly crucial in several applications, ranging from motion-driven virtual games to automated video surveillance systems. In these applications, sensors such as smart phone cameras, web cameras or CCTV cameras are used for detecting and tracking physical activities of users. Inevitably, spoof detection in HAR is essential to prevent anomalies and false alarms. To this end, we propose a deep learning based approach that can be used to detect spoofing in various fields such as border control, institutional security and public safety by surveillance cameras. Specifically, in this work, we address the problem of detecting spoofing occurring from video replay attacks, which is more common in such applications. We present a new database containing several videos of users juggling a football, captured under different lighting conditions and using different display and capture devices. We train our models using this database and the proposed system is capable of running in parallel with the HAR algorithms in real-time. Our experimental results show that our approach precisely detects video replay spoofing attacks and generalizes well, even to other applications such as spoof detection in face biometric authentication. Results show that our approach is effective even under resizing and compression artifacts that are common in HAR applications using remote server connections.
A Semi-Automatic Annotation Approach for Human Activity Recognition
