Face Identification and Tracking in Live Video Streams
The revolutionary motion-based face recognition technology enables video-based identification of human subjects with no prior enrolment. The API automatically recognizes all faces encountered in a video stream, registering their complete biometric information captured from the many different views and angles, complete with live emotions and expressions. Each subject can be tracked seamlessly and automatically without specific enrolment. Enrolling a person is as simple as putting a name tag in a video. This can be done at any time, and the system will automatically identify that subject in all past, present and future videos. Enrolment-free identification is perfect for building CRM systems for registration desks, access and attendance control systems, surveillance and security applications.
How does it differ from existing systems? Most current video identification systems are based on key frame processing. In other words, they discard information available in the motion stream, and revert to still image recognition instead. This design approach requires a set of sophisticated preliminary enrolment procedures where the subject's face is captured against plain background at various angles and posed expressions.
FaceInnovation implements a true motion-based video identification system. It automatically recognizes and tags all faces encountered in a video stream. Personalizing any identified subject becomes a simple matter of putting a name tag on it, or linking its tag to a database record. No special enrolment procedure is ever required. In addition, recognition rate is significantly higher in true motion-based recognition systems compared to traditional key frame-based ones. On comprehensive Multi-PIE tests, the recognition rate increases from 23% to 89% with video-based identification compared to key frame-based identification, considering that just a single image of a subject is enrolled and false acceptance rate is 0.06%.
The library can compare different faces, returning the degree of likeness. This allows identifying human faces appearing in still images or video streams by looking up face databases. Recognizing and identifying still images enables locating similar faces in driver's license databases while helping detect duplicates. The system implements image indexing, creating compact templates for faster searching. This in turn allows building a range of security applications such as video surveillance and real-time access control systems. Many more features and higher performance are achievable in video-based surveillance systems using the new set of motion-based recognition algorithms.
FaceInnovation is designed to perform equally well under varying lighting conditions. It works fine under daylight, fluorescent and incandescent lighting. When testing on a FRGC database, the library successfully identifies individuals in 93.9% of cases if acceptable false positive is 0.1%.
Face Detection in Stills and Videos
FaceInnovation returns coordinates of all human faces appearing in the picture – or notifies if no face is found. FaceInnovation can track all faces appearing in a video stream. It allows finding out if a new face appears in the frame, or if one of the subjects leaves the frame. This in turn enables easy implementation of people counting. When testing on the FERET dup1+gallery database (frontal passport-like photos) successfully detects 99.5% of faces, with only 0.05% false positives.