Ayonix has embedded its 3D face recognition technology into Axis IP camera(Camera model: Axis P1367). Axis camera has ARM9 with 512MB ram. ARM9 is the most frequently used CPU for cameras and embedded devices. The ARM is highly chosen for embedded platforms because of its fast architecture as well as low power consumption.
Ayonix R&D team worked hard to integrate face recognition into the IP camera for a long time. Camera resources are very limited, there is no GPU, CPU is 32bit single core. Therefore, Multicore 64bit Face recognition technology had to be redesigned from scratch. Redesigning engine for 32bit CPU limits Face recognition precision and increase FAR(False Acceptance rate) and FRR(False Rejection Rate)
Ayonix engineers worked hard to implement face detection, Face tracking, Face feature extraction and 3D face reconstruction. This was achieved but speed was very slow. So, some hardware tricks, assembly coding and some software primitives were developed specifically for the camera. Ayonix engineers refactored all vector computations inside single core ARM.
Face tracker was also another problem. Face tracker needs at least 30fps to track a face and compute face locations, face attributes(Gender, age, smile, eyes/mouth). In embedded platforms, hardware gives 20fps in 5MP images. Engineers developed a special algorithm to compensate missing frames.
Right now, camera runs face detection, face tracker, Gender and Age analysis, facial expression, face feature extraction, 3D face reconstruction and face matching. These run in 30fps in real-world conditions. Accuracy is exactly the same as Windows environment.
Ayonix’s solution can obtain 100ms recognition speed for any number of cameras (but the Ayonix server can handle 10 000 max.) and 550ms/1million records matching speed – regardless of the number of the cameras. The camera is available for system integrators, software companies as well as customers.
Source: www.ayonix.com; Dr. Sadi Vural