An Optimized Implementaion of Haar Like Feature Based Object Detection

Abstract

The AdaBoost (adaptive boosting) algorithm is widely used algorithm in computer vision and machine learning systems. It is a general method for generating a strong classifier out of a set of weak classifiers. The object detection algorithm by Viola and Jones [1] with Haar-like features as weak classifiers used AdaBoosting to construct a strong classifier cascade. This popular object detection algorithm runs in real time on desktop processors running in the range of 2GHz clock frequency. But the floating point arithmetic becomes a bottleneck for embedded and mobile platforms which has limited clock speeds for low power. This paper presents an optimized fixed point alternative to the floating point arithmetic used in the time critical classifier evaluation functions. The Open-CV library is used as the base software platform. The optimized fixed point implementation is tested with several test images consisting of one or more frontal faces. The results shows that the proposed implementation has a performance improvement from 3.81 to 5.84 fps.