Ehancing the Implementation of Adaboost Algorithm on a DSP-Based Platform

Adaboost algorithm for face detection in real-time is difficult to implement due to its significant computational complexity and memory requirements in terms of stringent memory bandwidth and huge memory volume. In this paper, we present two aspects of improvements in implementing the Adaboost algorithm, i.e. platform specific optimization on a DSP platform (TI´s TMS320DM642); and algorithm specific optimization including optimized cascade training, floating-point to fixed-point conversion (FFC) and scaling image. In the process of platform specific optimization, software pipeline, loop unrolling and writing liner assembly code is fulfilled. With these enhancements, we show in experimental results that the implemented system can detect human faces in real-time at a frame rate of 25 fps with little loss of correct detection rate. In our implementation, we further decrease the false detection rate, and dramatically reduce memory bandwidth and memory size required.