Universal Object Detection Architecture (UOD)


Phi Algorithm Solutions Inc. is developing a series of Object Detectors based on a common efficient architecture, suitable for a streamlined dataflow.  
This efficient architecture requires no preprocessing on subwindows, and therefore allows for a high degree of resource sharing.
By only changing coefficients within the architecture, any type of arbitrary high performance detector can be achieved.

FD-1 Face Detection
  • The FD-1 is a general face detector and first in a series of object detectors based on the UOD architecture. 
  • The detector is robust to a number of slight image distortions including in-plane rotation, out-of-plane rotations, translational motion, Gaussian noise, compression artefacts, poor contrast or non-uniform lighting, partial occlusions and oversmoothing.
  • The FD-1 detector attains a good balance of detection rates vs false alarms while remaining flexible with a high degree of programmability.

PD-1 Pedestrian Detection and Tracking
 
  • The PD-1 is a pedestrian detector and tracker that leverages the existing UOD architecture. 
  • The technology is a combination of three main functions: foreground extraction, pedestrian detection, and object tracking. The technology handles occlusions, and objects detected and tracked are assigned a unique label.
  • The PD-1 is robust to a number of slight image distortions and can be mapped to a number of implementations including CPU, DSP, RTL, and GPU.

Demo Pedestrian Detection and Tracking with CNN