Face Detection On The INCA General Information

Embedded systems such as mobile phones, handheld computer systems, robots etc., are becoming increasingly more important in our daily lives. There are countless examples of embedded systems on which it is feasible to have a robust face detector. For example to increase the interaction between a human and a robot, a first step for the robot could be to detect the face of the human it interacts with. However the majority of the available face detection techniques are designed to operate on high performance personal computer systems and are therefore usually not suited for embedded systems. The primary objective of our research was to find and implement a robust face detection technique, which is able to run at a reasonable speed, on an embedded system called the INCA. We implemented two of the most promising face detection techniques on the INCA: a skin color based face detector and the Viola and Jones face detector. We used a face detection experiment to determine which of these two techniques performed best. The experiment showed that the Viola and Jones detector performed best. Next we wanted to adapt the Viola and Jones detector so that it could run on the low-level SIMD processor which is part of the INCA. To increase the performance of the Viola and Jones detector, we propose two improvements of the algorithm. As a general improvement of the Viola and Jones algorithm, the usage of multiple threshold weak classifiers is proposed. A second improvement considers the usage of weak classifiers containing complex (Gabor) filters, instead of the simple Haarlike filters used in the original algorithm. Computing these complex filters would normally be computationally too expensive. We will demonstrate that when these complex filters are separable they can be computed rapidly on an SIMD processor.