Optimizing Image Processing on OMAP3 with driver-level frame buffering and color space conversion
Onboard image processing for unmanned aerial vehicles (UAVs) has become a popular method in the recent decades and the number of available hardware solutions has increased. Growing processing power and reduced weight and size of the embedded systems facilitate more computational power onboard the UAVs, making real-time image processing feasible. With respect to the software framework, the OpenCV library provides a set of useful functions to extract information from images. In this paper we first present a latency improvement over using OpenCV for camera input and show that the frame buffer optimization results in a latency reduction of up to one fifth compared to the OpenCV library. For the second part, we explain how utilizing direct driver access and hardware capabilities enables a faster color space conversion than OpenCV library functions. The color space conversion is tested with the L*a*b color space, which proves to be the right choice for our application, which is the detection of red objects in inhomogeneous light conditions. For our outdoor MAV application, the detection of six rectangular red objects takes no longer than about 50ms on average.
Johannes Schellen, Christian Dernehl and Stefan Kowalewski